diff --git a/apps/mobile/src/components/ProjectKnowledgeEntry.test.tsx b/apps/mobile/src/components/ProjectKnowledgeEntry.test.tsx
index dc628aa11f445..fea867f7cd530 100644
--- a/apps/mobile/src/components/ProjectKnowledgeEntry.test.tsx
+++ b/apps/mobile/src/components/ProjectKnowledgeEntry.test.tsx
@@ -245,7 +245,7 @@ const fact: ProjectFact = {
   tags: [],
   kind: "note",
   sources: [],
-  status: "proposed",
+  status: "saved",
   revision: 1,
   trust: 0.5,
   retrievalCount: 0,
@@ -486,7 +486,6 @@ it("shows user origin once without treating the assertion as a citation", async
   const asserted: ProjectFact = {
     ...fact,
     origin: { kind: "user" },
-    status: "confirmed",
     sources: [{ kind: "user" }],
   };
   state.list = { facts: [asserted], nextCursor: null };
@@ -681,7 +680,6 @@ it("marks a superseded fact read-only and lists superseded facts on request", as
   await screen.click(`Inspect fact: ${fact.content}`);
   const labels = [...screen.container.querySelectorAll("button")].map((item) => item.textContent);
   expect(labels).not.toContain("Edit");
-  expect(labels).not.toContain("Confirm");
   expect(labels).not.toContain("Supersede");
   await screen.click("Back");
   expect(state.listInputs.at(-1)).not.toHaveProperty("includeSuperseded");
@@ -841,7 +839,7 @@ it("drops a revised fact from the selection and the draft after saving", async (
   });
   screen.openMemory();
   await screen.click(`Inspect fact: ${fact.content}`);
-  await screen.click("Confirm");
+  await screen.click("Helpful");
   expect(state.setText).toHaveBeenCalledWith(scope.draftKey, "");
   expect(screen.container.textContent).toContain("0 selected");
   expect(screen.container.textContent).toContain("removed from the draft");
diff --git a/apps/mobile/src/components/ProjectKnowledgeEntry.tsx b/apps/mobile/src/components/ProjectKnowledgeEntry.tsx
index a664c574a46a3..4a0142b201da6 100644
--- a/apps/mobile/src/components/ProjectKnowledgeEntry.tsx
+++ b/apps/mobile/src/components/ProjectKnowledgeEntry.tsx
@@ -440,6 +440,8 @@ function ProjectKnowledgeSheet(
         <ScrollView
           className="flex-1"
           contentInsetAdjustmentBehavior="automatic"
+          automaticallyAdjustKeyboardInsets
+          keyboardDismissMode="on-drag"
           keyboardShouldPersistTaps="handled"
           contentContainerStyle={{ padding: 20, gap: 14 }}
         >
@@ -502,18 +504,17 @@ function ProjectKnowledgeSheet(
                   if (inspectedFact.status !== "forgotten")
                     push({ kind: "editor", mode: { kind, fact: inspectedFact } });
                 }}
-                onRevise={(action) => {
-                  const revise = () => void apply({ action, ...factMutation(inspectedFact) });
-                  if (action !== "forget") return revise();
+                onForget={() =>
                   Dialog.confirm({
                     title: "Forget this fact permanently?",
                     message:
                       "Its content, history and every copy sent to agents are removed and cannot be restored. If the fact changed rather than became wrong, supersede it instead.",
                     confirmText: "Forget",
                     destructive: true,
-                    onConfirm: revise,
-                  });
-                }}
+                    onConfirm: () =>
+                      void apply({ action: "forget", ...factMutation(inspectedFact) }),
+                  })
+                }
                 onFeedback={(helpful) =>
                   void apply({ action: "feedback", helpful, ...factMutation(inspectedFact) })
                 }
diff --git a/apps/mobile/src/components/ProjectKnowledgeFactDetail.tsx b/apps/mobile/src/components/ProjectKnowledgeFactDetail.tsx
index eb20cb7fd4448..76fb48a56374e 100644
--- a/apps/mobile/src/components/ProjectKnowledgeFactDetail.tsx
+++ b/apps/mobile/src/components/ProjectKnowledgeFactDetail.tsx
@@ -126,7 +126,7 @@ export function ProjectFactInspector(props: {
   readonly onReadSource: (ref: KnowledgeMessageRef) => void;
   readonly onOpenFact: OpenFact;
   readonly onEdit: (mode: "edit" | "supersede") => void;
-  readonly onRevise: (action: "confirm" | "forget") => void;
+  readonly onForget: () => void;
   readonly onFeedback: (helpful: boolean) => void;
   readonly onDismissHints: () => void;
 }) {
@@ -162,13 +162,6 @@ export function ProjectFactInspector(props: {
         <AppText className="text-xs text-foreground-muted">
           Trust signal {fact.trust.toFixed(2)} · revision {fact.revision}
         </AppText>
-        {fact.status === "proposed" &&
-        fact.origin.kind === "assisted" &&
-        fact.supersedes != null ? (
-          <AppText className="text-xs">
-            Confirming replaces fact {fact.supersedes.slice(0, 8)}; forgetting leaves it as it is.
-          </AppText>
-        ) : null}
       </View>
 
       <View className="gap-2">
@@ -202,14 +195,6 @@ export function ProjectFactInspector(props: {
 
       {active ? (
         <View className="flex-row flex-wrap gap-2">
-          {fact.status === "proposed" && !superseded ? (
-            <Button
-              label="Confirm"
-              tone="secondary"
-              disabled={!mutable}
-              onPress={() => props.onRevise("confirm")}
-            />
-          ) : null}
           {!superseded ? (
             <Button
               label="Edit"
@@ -238,12 +223,7 @@ export function ProjectFactInspector(props: {
             disabled={!mutable}
             onPress={() => props.onFeedback(false)}
           />
-          <Button
-            label="Forget"
-            tone="danger"
-            disabled={!mutable}
-            onPress={() => props.onRevise("forget")}
-          />
+          <Button label="Forget" tone="danger" disabled={!mutable} onPress={props.onForget} />
         </View>
       ) : null}
       {!props.canMutate && active ? (
@@ -358,9 +338,9 @@ export function KnowledgeHealthRow(props: {
   return (
     <View className="gap-1">
       <AppText className="text-xs text-foreground-muted">
-        {status.counts.confirmed} confirmed · {status.counts.proposed} proposed ·{" "}
-        {status.counts.superseded} superseded · {status.belowRecallThreshold} below recall ·{" "}
-        {status.feedback.helpful} helpful · {status.feedback.unhelpful} unhelpful
+        {status.counts.saved} saved · {status.counts.superseded} superseded ·{" "}
+        {status.belowRecallThreshold} below recall · {status.feedback.helpful} helpful ·{" "}
+        {status.feedback.unhelpful} unhelpful
       </AppText>
       {lost > 0 ? (
         <Button
diff --git a/apps/mobile/src/components/ProjectKnowledgeFactList.tsx b/apps/mobile/src/components/ProjectKnowledgeFactList.tsx
index 52dc97a77b607..1f7c4732ef5e1 100644
--- a/apps/mobile/src/components/ProjectKnowledgeFactList.tsx
+++ b/apps/mobile/src/components/ProjectKnowledgeFactList.tsx
@@ -35,7 +35,7 @@ import { KnowledgeHealthRow } from "./ProjectKnowledgeFactDetail";
 import { AppText, AppTextInput } from "./ui/AppText";
 import { Button } from "./ui/Button";
 
-type StatusFilter = "confirmed" | "proposed" | "suggested" | "forgotten";
+type StatusFilter = "saved" | "automatic" | "forgotten";
 
 /** Filters, search and paging for the fact list. The sheet reads its results for draft labels. */
 export function useProjectKnowledgeBrowse(props: {
@@ -44,7 +44,7 @@ export function useProjectKnowledgeBrowse(props: {
   readonly history: boolean;
   readonly onArchivedChange: () => void;
 }) {
-  const [status, setStatus] = useState<StatusFilter>("confirmed");
+  const [status, setStatus] = useState<StatusFilter>("saved");
   const [afterId, setAfterId] = useState<ProjectFactId | undefined>();
   const [includeArchived, setIncludeArchived] = useState(false);
   const [searchText, setSearchText] = useState("");
@@ -56,11 +56,10 @@ export function useProjectKnowledgeBrowse(props: {
       environmentId: props.environmentId,
       input: {
         projectId: props.projectId,
-        // The lost-source filter spans proposed and confirmed, matching the health count.
         ...(props.history && lostSourcesOnly
           ? { sourceState: "unavailable" as const }
-          : status === "suggested"
-            ? { status: "proposed" as const, origin: "assisted" as const }
+          : status === "automatic"
+            ? { status: "saved" as const, origin: "assisted" as const }
             : { status }),
         ...(afterId ? { afterId } : {}),
         ...(props.history && showSuperseded ? { includeSuperseded: true } : {}),
@@ -214,7 +213,6 @@ export function ProjectKnowledgeFactList(props: {
 }) {
   const { browse } = props;
   const { list, health, search, searchInput } = browse;
-  const suggested = health.data?.assistedLearning?.suggested;
   return (
     <View className="gap-4">
       <View className="gap-2">
@@ -243,12 +241,12 @@ export function ProjectKnowledgeFactList(props: {
 
       <View className="flex-row flex-wrap gap-2">
         {(props.assisted
-          ? (["confirmed", "proposed", "suggested", "forgotten"] as const)
-          : (["confirmed", "proposed", "forgotten"] as const)
+          ? (["saved", "automatic", "forgotten"] as const)
+          : (["saved", "forgotten"] as const)
         ).map((value) => (
           <Button
             key={value}
-            label={value === "suggested" && suggested ? `suggested (${suggested})` : value}
+            label={value}
             tone={
               browse.status === value && !searchInput && !browse.lostSourcesOnly
                 ? "primary"
@@ -427,7 +425,6 @@ function FactRow(props: {
   );
 }
 
-/** The project's opt-in for model-suggested facts and what the last run did with it. */
 function AssistedLearningRow(props: {
   readonly environmentId: EnvironmentId;
   readonly projectId: ProjectId;
@@ -463,17 +460,17 @@ function AssistedLearningRow(props: {
   return (
     <View className="gap-1 rounded-2xl bg-subtle p-3">
       <View className="flex-row items-center gap-3">
-        <AppText className="flex-1">Suggest facts from finished runs</AppText>
+        <AppText className="flex-1">Save facts from finished runs</AppText>
         <Switch
-          accessibilityLabel="Suggest facts from finished runs"
+          accessibilityLabel="Save facts from finished runs"
           value={enabled}
           disabled={!canWrite || busy}
           onValueChange={(value) => void toggle(value)}
         />
       </View>
       <AppText className="text-xs text-foreground-muted">
-        After each run in this project, your text generation model proposes up to three cited facts
-        for review. Each run costs one model call.
+        After each run in this project, your text generation model saves up to three cited facts.
+        Each run costs one model call.
       </AppText>
       {props.lastRun ? (
         <AppText className="text-xs">{assistedLearningRunLabel(props.lastRun)}</AppText>
diff --git a/apps/mobile/src/features/settings/SettingsServerControlsRouteScreen.test.tsx b/apps/mobile/src/features/settings/SettingsServerControlsRouteScreen.test.tsx
index a58f2991dee99..c4377efd1e523 100644
--- a/apps/mobile/src/features/settings/SettingsServerControlsRouteScreen.test.tsx
+++ b/apps/mobile/src/features/settings/SettingsServerControlsRouteScreen.test.tsx
@@ -212,13 +212,13 @@ describe("SettingsServerControlsRouteScreen", () => {
       act(() => root.render(<SettingsZackRouteScreen />));
       const toggle = Array.from(
         container.querySelectorAll<HTMLButtonElement>('[role="switch"]'),
-      ).find((button) => button.textContent === "Suggest facts from finished runs")!;
+      ).find((button) => button.textContent === "Save facts from finished runs")!;
 
       act(() => toggle.click());
 
       expect(toggle.disabled).toBe(true);
       expect(mocks.updateSettings).not.toHaveBeenCalled();
-      expect(container.textContent).toContain("Assisted learning is not supported on Old server.");
+      expect(container.textContent).toContain("Automatic memory is not supported on Old server.");
     },
   );
 
@@ -228,33 +228,35 @@ describe("SettingsServerControlsRouteScreen", () => {
     act(() => root.render(<SettingsZackRouteScreen />));
     const toggle = Array.from(
       container.querySelectorAll<HTMLButtonElement>('[role="switch"]'),
-    ).find((button) => button.textContent === "Suggest facts from finished runs")!;
+    ).find((button) => button.textContent === "Save facts from finished runs")!;
 
     act(() => toggle.click());
 
     expect(toggle.disabled).toBe(true);
     expect(mocks.updateSettings).not.toHaveBeenCalled();
     expect(container.textContent).toContain(
-      "Assisted learning is supported on only some selected environments.",
+      "Automatic memory is supported on only some selected environments.",
     );
     expect(container.textContent).toContain("Old server");
   });
 
-  it("enables assisted learning for a supported project without targeting unrelated older environments", async () => {
+  it("turns automatic memory off for a supported project without targeting unrelated older environments", async () => {
     mocks.targets = [environment("new", "New server", true), environment("old", "Old server")];
     mocks.projectSelected = true;
     mocks.projectEnvironmentIds = ["new"];
     act(() => root.render(<SettingsZackRouteScreen />));
     const toggle = Array.from(
       container.querySelectorAll<HTMLButtonElement>('[role="switch"]'),
-    ).find((button) => button.textContent === "Suggest facts from finished runs")!;
+    ).find((button) => button.textContent === "Save facts from finished runs")!;
 
     await act(async () => toggle.click());
 
     expect(toggle.disabled).toBe(false);
     expect(mocks.updateSettings).toHaveBeenCalledExactlyOnceWith({
       environmentId: "new",
-      input: { patch: { projectSettingsOverrides: { project: { enableAssistedLearning: true } } } },
+      input: {
+        patch: { projectSettingsOverrides: { project: { enableAssistedLearning: false } } },
+      },
     });
   });
 
diff --git a/apps/mobile/src/features/settings/SettingsServerControlsRouteScreen.tsx b/apps/mobile/src/features/settings/SettingsServerControlsRouteScreen.tsx
index 595c083638c6d..44cef66bd3c4b 100644
--- a/apps/mobile/src/features/settings/SettingsServerControlsRouteScreen.tsx
+++ b/apps/mobile/src/features/settings/SettingsServerControlsRouteScreen.tsx
@@ -202,8 +202,8 @@ function ServerSettingsDetail(props: { readonly page: SettingsPage }) {
     unsupportedAssistedLearning.length === 0
       ? null
       : unsupportedAssistedLearning.length < targets.length
-        ? `Assisted learning is supported on only some selected environments. Update T3 Code on ${unsupportedAssistedLearning.map((target) => target.environment.label).join(", ")}, or select a supported environment.`
-        : `Assisted learning is not supported on ${unsupportedAssistedLearning.map((target) => target.environment.label).join(", ")}. Update T3 Code on ${unsupportedAssistedLearning.length === 1 ? "that environment" : "those environments"}.`;
+        ? `Automatic memory is supported on only some selected environments. Update T3 Code on ${unsupportedAssistedLearning.map((target) => target.environment.label).join(", ")}, or select a supported environment.`
+        : `Automatic memory is not supported on ${unsupportedAssistedLearning.map((target) => target.environment.label).join(", ")}. Update T3 Code on ${unsupportedAssistedLearning.length === 1 ? "that environment" : "those environments"}.`;
   const [pendingWrites, setPendingWrites] = useState(0);
   const writeInFlight = useRef(false);
   // Closing views can save after controls disable; queue those patches behind the pending write.
@@ -424,10 +424,10 @@ function ServerSettingsDetail(props: { readonly page: SettingsPage }) {
                 <SettingsSection title="Project memory">
                   <SettingsSwitchRow
                     icon="brain"
-                    label="Suggest facts from finished runs"
+                    label="Save facts from finished runs"
                     subtitle={
                       assistedLearningUnavailableReason ??
-                      "After each run, your text generation model proposes up to three cited facts for review. Each run costs one model call."
+                      "After each run, your text generation model saves up to three cited facts. Each run costs one model call."
                     }
                     value={uniform("enableAssistedLearning")}
                     disabled={
diff --git a/apps/server/src/mcp/toolkits/knowledge/tools.ts b/apps/server/src/mcp/toolkits/knowledge/tools.ts
index 234b4bbebdf46..3ac8b512616cd 100644
--- a/apps/server/src/mcp/toolkits/knowledge/tools.ts
+++ b/apps/server/src/mcp/toolkits/knowledge/tools.ts
@@ -47,7 +47,7 @@ export const KnowledgeToolkit = Toolkit.make(
   Tool.make("t3_knowledge_list", {
     ...shared,
     description:
-      "List project facts, including proposed or forgotten facts when requested. Superseded facts are excluded unless includeSuperseded is set; sourceState unavailable lists only facts whose citations no longer resolve. This works independently of the retrieval kernel.",
+      "List saved project facts, or forgotten facts when requested. Superseded facts are excluded unless includeSuperseded is set; sourceState unavailable lists only facts whose citations no longer resolve. This works independently of the retrieval kernel.",
     parameters: Schema.Struct({
       ...ProjectKnowledgeListInput.fields,
       environmentId: OrchestratorMcpEnvironmentTarget,
@@ -59,7 +59,7 @@ export const KnowledgeToolkit = Toolkit.make(
   Tool.make("t3_knowledge_read_fact", {
     ...shared,
     description:
-      "Read one project fact with its recent revision, confirmation, feedback and supersession history, linked superseding facts, and possible duplicates or conflicts found by lexical similarity.",
+      "Read one project fact with its recent revision, feedback and supersession history, linked superseding facts, and possible duplicates or conflicts found by lexical similarity.",
     parameters: Schema.Struct({
       ...ProjectKnowledgeReadFactInput.fields,
       environmentId: OrchestratorMcpEnvironmentTarget,
@@ -107,7 +107,7 @@ export const KnowledgeToolkit = Toolkit.make(
   Tool.make("t3_knowledge_mutate", {
     ...shared,
     description:
-      "Propose or edit project knowledge, supersede a fact with a replacement, forget a fact, or give trust feedback. When a fact comes from a conversation message, attach that exact message citation. Leave sources empty only for an agent observation with no message source; its origin is recorded from this authenticated caller. Proposed facts can be recalled immediately, but are not automatically added to the prompt. Confirmation in the app is optional and remains a separate trust action. Agents cannot confirm facts or claim direct user assertions. Reuse commandId when retrying. Updates require the current revision. Supersede creates replacementFactId and keeps the old fact's content but removes it from recall; prefer it over forget when knowledge changed. Created, edited and replacement facts return hints naming possible duplicates or conflicts to review. Forgotten facts cannot be restored.",
+      "Create or edit project knowledge, supersede a fact with a replacement, forget a fact, or give trust feedback. When a fact comes from a conversation message, attach that exact message citation. Leave sources empty only for an agent observation with no message source; its origin is recorded from this authenticated caller. Facts are saved immediately and can be recalled by search, but are not automatically added to the prompt. Saving records a claim with its origin, not proof that it is true. Agents cannot claim direct user assertions. Reuse commandId when retrying. Updates require the current revision. Supersede creates replacementFactId and keeps the old fact's content but removes it from recall; prefer it over forget when knowledge changed. Created, edited and replacement facts return hints naming possible duplicates or conflicts to review. Forgotten facts cannot be restored.",
     parameters: Schema.Struct({
       mutation: ProjectKnowledgeMutation,
       environmentId: OrchestratorMcpEnvironmentTarget,
diff --git a/apps/server/src/orchestration-v2/ProviderRuntimeRecoveryService.test.ts b/apps/server/src/orchestration-v2/ProviderRuntimeRecoveryService.test.ts
index 957165e4674f5..fe353fdd4088b 100644
--- a/apps/server/src/orchestration-v2/ProviderRuntimeRecoveryService.test.ts
+++ b/apps/server/src/orchestration-v2/ProviderRuntimeRecoveryService.test.ts
@@ -654,7 +654,7 @@ it.effect.each([
     turnItems: [],
   } as unknown as OrchestrationV2ThreadProjection;
   const layer = ProviderRuntimeRecovery.layer.pipe(
-    Layer.provide(ServerSettings.layerTest()),
+    Layer.provide(ServerSettings.layerTest({ enableAssistedLearning: false })),
     Layer.provide(
       Layer.mergeAll(
         Layer.mock(ProjectionStore.ProjectionStoreV2)({
diff --git a/apps/server/src/orchestration-v2/RestartContinuation.test.ts b/apps/server/src/orchestration-v2/RestartContinuation.test.ts
index a318130500b3b..118fbf0719dde 100644
--- a/apps/server/src/orchestration-v2/RestartContinuation.test.ts
+++ b/apps/server/src/orchestration-v2/RestartContinuation.test.ts
@@ -250,6 +250,7 @@ it.effect.each([
           Layer.mergeAll(
             ServerSettings.layerTest({
               continueThreadsAfterServerUpdate: enabled,
+              enableAssistedLearning: false,
               projectSettingsOverrides:
                 projectOverride === undefined
                   ? {}
@@ -386,7 +387,10 @@ it.effect.each(["running", "waiting"])(
       const recovery = yield* ProviderRuntimeRecovery.make.pipe(
         Effect.provide(
           Layer.mergeAll(
-            ServerSettings.layerTest({ continueThreadsAfterServerUpdate: true }),
+            ServerSettings.layerTest({
+              continueThreadsAfterServerUpdate: true,
+              enableAssistedLearning: false,
+            }),
             Layer.mock(ProjectionStore.ProjectionStoreV2)({
               getRecoveryThreadIds: () => Effect.succeed([threadId]),
               getRuntimeRecoveryProjection: () => Effect.succeed(projection),
@@ -633,7 +637,10 @@ it.effect.each([
         const recovery = yield* ProviderRuntimeRecovery.make.pipe(
           Effect.provide(
             Layer.mergeAll(
-              ServerSettings.layerTest({ continueThreadsAfterServerUpdate: true }),
+              ServerSettings.layerTest({
+                continueThreadsAfterServerUpdate: true,
+                enableAssistedLearning: false,
+              }),
               Layer.mock(ProjectionStore.ProjectionStoreV2)({
                 getRecoveryThreadIds: () => Effect.succeed([threadId]),
                 getRuntimeRecoveryProjection: () => Effect.succeed(projection),
diff --git a/apps/server/src/orchestration-v2/RunExecutionService.test.ts b/apps/server/src/orchestration-v2/RunExecutionService.test.ts
index a4cd9375c6810..406927b1d4daf 100644
--- a/apps/server/src/orchestration-v2/RunExecutionService.test.ts
+++ b/apps/server/src/orchestration-v2/RunExecutionService.test.ts
@@ -3259,6 +3259,7 @@ function captureRootRunTermination(input: {
             ingestNormalized: input.ingestNormalized ?? (() => Effect.succeed([])),
           }),
           ServerSettings.layerTest({
+            enableAssistedLearning: false,
             enableModelEvaluations: input.enableModelEvaluations ?? false,
           }),
           Layer.succeed(RunFinalizationService.RunFinalizationObserver, {
diff --git a/apps/server/src/persistence/Migrations.test.ts b/apps/server/src/persistence/Migrations.test.ts
index 733b0258de1c4..9787393005f67 100644
--- a/apps/server/src/persistence/Migrations.test.ts
+++ b/apps/server/src/persistence/Migrations.test.ts
@@ -8,6 +8,8 @@ import * as Path from "effect/Path";
 import * as SqlClient from "effect/sql/SqlClient";
 
 import { migrationManifest, runMigrations } from "./Migrations.ts";
+import ProjectLoomTimeIndexes from "./Migrations/075_ProjectLoomTimeIndexes.ts";
+import ProjectLoomBoundedReads from "./Migrations/077_ProjectLoomBoundedReads.ts";
 
 const v2Migrations = migrationManifest.filter(([id]) => id >= 55);
 const newerMigrationId = Math.max(...migrationManifest.map(([id]) => id)) + 1;
@@ -23,6 +25,7 @@ const modelEvaluationMigration = [88, "ModelEvaluations"] as const;
 const delegatedRecoveryMigration = [89, "DelegatedRecoveryIndexes"] as const;
 const scheduledTaskThreadModeMigration = [90, "ScheduledTaskThreadMode"] as const;
 const dropProjectLoomIndexesMigration = [91, "DropProjectLoomIndexes"] as const;
+const savedLifecycleMigration = [92, "ProjectKnowledgeSavedLifecycle"] as const;
 
 it.effect("upgrades attempt receipts atomically and preserves distinct steering messages", () =>
   Effect.gen(function* () {
@@ -49,6 +52,7 @@ it.effect("upgrades attempt receipts atomically and preserves distinct steering
       delegatedRecoveryMigration,
       scheduledTaskThreadModeMigration,
       dropProjectLoomIndexesMigration,
+      savedLifecycleMigration,
     ]);
     assert.deepStrictEqual(yield* sql`SELECT * FROM project_knowledge_deliveries`, before);
     yield* sql`INSERT INTO project_knowledge_deliveries (
@@ -67,6 +71,281 @@ it.effect("upgrades attempt receipts atomically and preserves distinct steering
   }).pipe(Effect.provide(NodeSqliteClient.layer({ filename: ":memory:" }))),
 );
 
+it.effect("rebuilds knowledge facts as saved or forgotten, atomically and in place", () =>
+  Effect.gen(function* () {
+    const sql = yield* SqlClient.SqlClient;
+    yield* runMigrations({ toMigrationInclusive: 91 });
+    const insertFact = (fact: {
+      readonly rowid: number;
+      readonly id: string;
+      readonly status: string;
+      readonly origin: string;
+      readonly projectId?: string;
+      readonly supersedes?: string;
+      readonly supersededBy?: string;
+    }) => {
+      const forgotten = fact.status === "forgotten";
+      return sql`
+        INSERT INTO project_knowledge_facts (
+          rowid, id, project_id, status, revision, content, entities_json, tags_json, kind, trust,
+          retrieval_count, created_at, updated_at, forgotten_at, sources_json, origin_json,
+          supersedes_id, superseded_by_id
+        ) VALUES (
+          ${fact.rowid}, ${fact.id}, ${fact.projectId ?? "project"}, ${fact.status},
+          ${fact.rowid + 1}, ${forgotten ? null : `lifecycle word${fact.id}`}, '[]', '[]',
+          'decision', 0.65, ${fact.rowid * 2}, '2026-01-01T00:00:00.000Z',
+          '2026-02-02T00:00:00.000Z', ${forgotten ? "2026-03-03T00:00:00.000Z" : null},
+          ${forgotten ? "[]" : '[{"kind":"user"}]'}, ${fact.origin},
+          ${fact.supersedes ?? null}, ${fact.supersededBy ?? null}
+        )
+      `;
+    };
+    const user = '{"kind":"user"}';
+    const agent = '{"kind":"agent-client","sessionId":"session","label":"Agent"}';
+    const assisted = '{"kind":"assisted","threadId":"thread","runId":"run","runOutcome":"failed"}';
+    yield* insertFact({
+      rowid: 3,
+      id: "replaced",
+      status: "confirmed",
+      origin: user,
+      supersededBy: "replacer",
+    });
+    yield* insertFact({
+      rowid: 7,
+      id: "replacer",
+      status: "proposed",
+      origin: assisted,
+      supersedes: "replaced",
+    });
+    yield* insertFact({
+      rowid: 11,
+      id: "edited-pending",
+      status: "proposed",
+      origin: user,
+      supersedes: "target",
+    });
+    yield* insertFact({
+      rowid: 15,
+      id: "assisted-pending",
+      status: "proposed",
+      origin: assisted,
+      supersedes: "target",
+    });
+    yield* insertFact({ rowid: 19, id: "target", status: "confirmed", origin: agent });
+    yield* insertFact({
+      rowid: 23,
+      id: "dangling",
+      status: "confirmed",
+      origin: agent,
+      supersedes: "missing",
+    });
+    yield* insertFact({
+      rowid: 27,
+      id: "forgotten-pending",
+      status: "forgotten",
+      origin: agent,
+      supersedes: "target",
+    });
+    yield* insertFact({
+      rowid: 31,
+      id: "cross-project",
+      status: "proposed",
+      origin: agent,
+      projectId: "other",
+      supersedes: "foreign",
+    });
+    yield* insertFact({
+      rowid: 35,
+      id: "foreign",
+      status: "confirmed",
+      origin: user,
+      supersededBy: "cross-project",
+    });
+    yield* sql`
+      INSERT INTO project_knowledge_fact_history (
+        project_id, fact_id, revision, event, actor_json, sources_json, trust_after, created_at
+      ) VALUES ('project', 'target', 20, 'confirmed', ${user}, '[]', 0.65, '2026-02-02T00:00:00.000Z')
+    `;
+    yield* sql`
+      INSERT INTO project_knowledge_commands (
+        command_id, project_id, fact_id, request_digest, search_json
+      ) VALUES ('command', 'project', 'target', 'digest', '{"rootRequestId":"command"}')
+    `;
+    yield* sql`
+      INSERT INTO project_knowledge_assisted_runs (
+        run_id, project_id, thread_id, run_outcome, result, reason, proposed, suggested, completed_at
+      ) VALUES ('run', 'project', 'thread', 'failed', 'ran', 'kept', 2, 1, '2026-02-02T00:00:00.000Z')
+    `;
+    const facts = sql<Record<string, unknown>>`
+      SELECT rowid, * FROM project_knowledge_facts ORDER BY rowid
+    `;
+    const history = sql`SELECT * FROM project_knowledge_fact_history ORDER BY sequence`;
+    const commands = sql`SELECT * FROM project_knowledge_commands ORDER BY command_id`;
+    const before = yield* facts;
+    const historyBefore = yield* history;
+    const commandsBefore = yield* commands;
+    const schema = sql`SELECT name, sql FROM sqlite_master ORDER BY name`;
+    const schemaBefore = yield* schema;
+
+    yield* sql`CREATE VIEW project_knowledge_facts_saved AS SELECT 1`;
+    const failedSchema = yield* schema;
+    assert.isTrue(Exit.isFailure(yield* Effect.exit(runMigrations())));
+    assert.deepStrictEqual(yield* facts, before);
+    assert.deepStrictEqual(yield* schema, failedSchema);
+    yield* sql`DROP VIEW project_knowledge_facts_saved`;
+    assert.deepStrictEqual(yield* schema, schemaBefore);
+
+    assert.deepStrictEqual(yield* runMigrations(), [savedLifecycleMigration]);
+
+    const cleared = new Set([
+      "edited-pending",
+      "assisted-pending",
+      "dangling",
+      "forgotten-pending",
+      "cross-project",
+    ]);
+    assert.deepStrictEqual(
+      yield* facts,
+      before.map((row) => ({
+        ...row,
+        status: row.status === "forgotten" ? "forgotten" : "saved",
+        supersedes_id: cleared.has(String(row.id)) ? null : row.supersedes_id,
+      })),
+    );
+    assert.deepStrictEqual(yield* history, historyBefore);
+    assert.deepStrictEqual(yield* commands, commandsBefore);
+    assert.deepStrictEqual(yield* sql`SELECT * FROM project_knowledge_assisted_runs`, [
+      {
+        run_id: "run",
+        project_id: "project",
+        thread_id: "thread",
+        run_outcome: "failed",
+        result: "ran",
+        reason: "kept",
+        saved: 2,
+        completed_at: "2026-02-02T00:00:00.000Z",
+      },
+    ]);
+    assert.deepStrictEqual(
+      yield* sql`
+        SELECT name FROM sqlite_master
+        WHERE tbl_name = 'project_knowledge_facts' AND type IN ('index', 'trigger')
+          AND name NOT LIKE 'sqlite_autoindex%'
+        ORDER BY name
+      `,
+      [
+        { name: "project_knowledge_delete" },
+        { name: "project_knowledge_facts_project" },
+        { name: "project_knowledge_insert" },
+        { name: "project_knowledge_update" },
+      ],
+    );
+
+    const matches = (term: string) =>
+      sql<{ id: string }>`
+        SELECT f.id FROM project_knowledge_fts
+        JOIN project_knowledge_facts f ON f.rowid = project_knowledge_fts.rowid
+        WHERE project_knowledge_fts MATCH ${term} ORDER BY f.id
+      `.pipe(Effect.map((rows) => rows.map((row) => row.id)));
+    assert.deepStrictEqual(yield* matches("wordtarget"), ["target"]);
+    assert.lengthOf(yield* matches("lifecycle"), 8);
+    yield* sql`UPDATE project_knowledge_facts SET content = 'renamed' WHERE id = 'target'`;
+    assert.deepStrictEqual(yield* matches("wordtarget"), []);
+    assert.deepStrictEqual(yield* matches("renamed"), ["target"]);
+    yield* sql`DELETE FROM project_knowledge_facts WHERE id = 'target'`;
+    assert.deepStrictEqual(yield* matches("renamed"), []);
+    yield* insertFact({ rowid: 40, id: "fresh", status: "saved", origin: user });
+    assert.deepStrictEqual(yield* matches("wordfresh"), ["fresh"]);
+
+    for (const status of ["proposed", "confirmed"]) {
+      assert.isTrue(
+        Exit.isFailure(
+          yield* Effect.exit(insertFact({ rowid: 50, id: `old-${status}`, status, origin: user })),
+        ),
+      );
+      assert.isTrue(
+        Exit.isFailure(
+          yield* Effect.exit(
+            sql`UPDATE project_knowledge_facts SET status = ${status} WHERE id = 'fresh'`,
+          ),
+        ),
+      );
+    }
+    assert.isTrue(
+      Exit.isFailure(
+        yield* Effect.exit(
+          sql`UPDATE project_knowledge_assisted_runs SET saved = -1 WHERE run_id = 'run'`,
+        ),
+      ),
+    );
+    assert.deepStrictEqual(yield* runMigrations(), []);
+
+    // The published preview recorded this same saved schema at 91, before Loom cleanup.
+    const savedFacts = yield* facts;
+    const savedHistory = yield* history;
+    const savedCommands = yield* commands;
+    const savedRuns = yield* sql`SELECT * FROM project_knowledge_assisted_runs`;
+    const savedSchema = yield* schema;
+    yield* ProjectLoomTimeIndexes;
+    yield* ProjectLoomBoundedReads;
+    yield* previewLoomIndexes;
+    yield* sql`DELETE FROM effect_sql_migrations WHERE migration_id = 91`;
+    yield* sql`UPDATE effect_sql_migrations SET migration_id = 91 WHERE migration_id = 92`;
+    const migrations = sql<{
+      readonly migration_id: number;
+      readonly name: string;
+      readonly created_at: string;
+    }>`SELECT migration_id, name, created_at FROM effect_sql_migrations ORDER BY migration_id`;
+    const previewHistory = yield* migrations;
+    const previewSchema = yield* schema;
+
+    yield* sql`INSERT INTO effect_sql_migrations (migration_id, name) VALUES (93, 'UnknownPreview')`;
+    const unknownHistory = yield* migrations;
+    assert.isTrue(Exit.isFailure(yield* Effect.exit(runMigrations())));
+    assert.deepStrictEqual(yield* migrations, unknownHistory);
+    assert.deepStrictEqual(yield* schema, previewSchema);
+    yield* sql`DELETE FROM effect_sql_migrations WHERE migration_id = 93`;
+
+    yield* sql`ALTER TABLE project_knowledge_assisted_runs RENAME COLUMN saved TO proposed`;
+    const incompleteSchema = yield* schema;
+    assert.isTrue(Exit.isFailure(yield* Effect.exit(runMigrations())));
+    assert.deepStrictEqual(yield* migrations, previewHistory);
+    assert.deepStrictEqual(yield* schema, incompleteSchema);
+    yield* sql`ALTER TABLE project_knowledge_assisted_runs RENAME COLUMN proposed TO saved`;
+
+    yield* sql`
+      CREATE TRIGGER reject_saved_preview_receipt BEFORE UPDATE ON effect_sql_migrations
+      WHEN new.migration_id = 92 BEGIN
+        SELECT RAISE(ABORT, 'receipt update failed');
+      END
+    `;
+    const rollbackSchema = yield* schema;
+    assert.isTrue(Exit.isFailure(yield* Effect.exit(runMigrations())));
+    assert.deepStrictEqual(yield* migrations, previewHistory);
+    assert.deepStrictEqual(yield* schema, rollbackSchema);
+    yield* sql`DROP TRIGGER reject_saved_preview_receipt`;
+
+    assert.deepStrictEqual(yield* runMigrations(), [dropProjectLoomIndexesMigration]);
+    assert.deepStrictEqual(yield* facts, savedFacts);
+    assert.deepStrictEqual(yield* history, savedHistory);
+    assert.deepStrictEqual(yield* commands, savedCommands);
+    assert.deepStrictEqual(yield* sql`SELECT * FROM project_knowledge_assisted_runs`, savedRuns);
+    assert.deepStrictEqual(yield* schema, savedSchema);
+    const reconciledHistory = yield* migrations;
+    assert.deepStrictEqual(
+      reconciledHistory.map(({ migration_id, name }) => [migration_id, name] as const),
+      migrationManifest,
+    );
+    assert.deepStrictEqual(
+      reconciledHistory.find(({ migration_id }) => migration_id === 92),
+      { ...previewHistory.find(({ migration_id }) => migration_id === 91), migration_id: 92 },
+    );
+    assert.deepStrictEqual(yield* matches("wordfresh"), ["fresh"]);
+    assert.deepStrictEqual(yield* runMigrations(), []);
+    assert.deepStrictEqual(yield* migrations, reconciledHistory);
+  }).pipe(Effect.provide(NodeSqliteClient.layer({ filename: ":memory:" }))),
+);
+
 const previewNativeColumns = Effect.gen(function* () {
   const sql = yield* SqlClient.SqlClient;
   yield* sql`ALTER TABLE project_knowledge_deliveries ADD COLUMN provider_thread_id TEXT`;
@@ -396,6 +675,7 @@ describe("runMigrations", () => {
             delegatedRecoveryMigration,
             scheduledTaskThreadModeMigration,
             dropProjectLoomIndexesMigration,
+            savedLifecycleMigration,
           ],
         );
         yield* assertConsolidatedSchema;
@@ -438,6 +718,7 @@ describe("runMigrations", () => {
           delegatedRecoveryMigration,
           scheduledTaskThreadModeMigration,
           dropProjectLoomIndexesMigration,
+          savedLifecycleMigration,
         ],
       );
       assert.deepStrictEqual(yield* runMigrations(), [
@@ -453,6 +734,7 @@ describe("runMigrations", () => {
         delegatedRecoveryMigration,
         scheduledTaskThreadModeMigration,
         dropProjectLoomIndexesMigration,
+        savedLifecycleMigration,
       ]);
       yield* assertConsolidatedSchema;
       assert.deepStrictEqual(
@@ -490,6 +772,7 @@ describe("runMigrations", () => {
           delegatedRecoveryMigration,
           scheduledTaskThreadModeMigration,
           dropProjectLoomIndexesMigration,
+          savedLifecycleMigration,
         ]);
 
         yield* assertConsolidatedSchema;
@@ -597,6 +880,7 @@ describe("runMigrations", () => {
         delegatedRecoveryMigration,
         scheduledTaskThreadModeMigration,
         dropProjectLoomIndexesMigration,
+        savedLifecycleMigration,
       ]);
       yield* assertConsolidatedSchema;
       assert.deepStrictEqual(yield* sql`SELECT secret_name, fenced FROM peer_environment_links`, [
@@ -653,6 +937,7 @@ describe("runMigrations", () => {
           delegatedRecoveryMigration,
           scheduledTaskThreadModeMigration,
           dropProjectLoomIndexesMigration,
+          savedLifecycleMigration,
         ]);
         assert.deepStrictEqual(yield* history, migrationManifest);
         assert.deepStrictEqual(
@@ -714,6 +999,7 @@ describe("runMigrations", () => {
           delegatedRecoveryMigration,
           scheduledTaskThreadModeMigration,
           dropProjectLoomIndexesMigration,
+          savedLifecycleMigration,
         ]);
         assert.deepStrictEqual(yield* history, migrationManifest);
         assert.deepStrictEqual(yield* sql`SELECT * FROM cross_environment_launches`, routes);
diff --git a/apps/server/src/persistence/Migrations.ts b/apps/server/src/persistence/Migrations.ts
index 907a052cea6c4..ba40649f0145c 100644
--- a/apps/server/src/persistence/Migrations.ts
+++ b/apps/server/src/persistence/Migrations.ts
@@ -3,6 +3,7 @@ import * as Effect from "effect/Effect";
 import * as SqlClient from "effect/sql/SqlClient";
 import { reconcileV2PreviewMigration } from "./reconcileV2PreviewMigration.ts";
 import { reconcileProjectKnowledgePreviewMigration } from "./reconcileProjectKnowledgePreviewMigration.ts";
+import { reconcileSavedKnowledgePreviewMigration } from "./reconcileSavedKnowledgePreviewMigration.ts";
 
 import Migration0001 from "./Migrations/001_OrchestrationEvents.ts";
 import Migration0002 from "./Migrations/002_OrchestrationCommandReceipts.ts";
@@ -96,6 +97,7 @@ import Migration0088 from "./Migrations/088_ModelEvaluations.ts";
 import Migration0089 from "./Migrations/089_DelegatedRecoveryIndexes.ts";
 import Migration0090 from "./Migrations/090_ScheduledTaskThreadMode.ts";
 import Migration0091 from "./Migrations/091_DropProjectLoomIndexes.ts";
+import Migration0092 from "./Migrations/092_ProjectKnowledgeSavedLifecycle.ts";
 
 const migrationEntries = [
   [1, "OrchestrationEvents", Migration0001],
@@ -190,6 +192,7 @@ const migrationEntries = [
   [89, "DelegatedRecoveryIndexes", Migration0089],
   [90, "ScheduledTaskThreadMode", Migration0090],
   [91, "DropProjectLoomIndexes", Migration0091],
+  [92, "ProjectKnowledgeSavedLifecycle", Migration0092],
 ] as const;
 
 export const migrationManifest = migrationEntries.map(([id, name]) => [id, name] as const);
@@ -254,12 +257,17 @@ export const runMigrations = Effect.fn("runMigrations")(function* ({
     toMigrationInclusive === undefined || toMigrationInclusive >= 80
       ? yield* reconcileProjectKnowledgePreviewMigration(migrationManifest)
       : [];
+  const savedKnowledgePreviewMigrations =
+    toMigrationInclusive === undefined || toMigrationInclusive >= 92
+      ? yield* reconcileSavedKnowledgePreviewMigration(migrationManifest)
+      : [];
   if (includesV2) {
     yield* assertV2MigrationsCanRun(toMigrationInclusive);
   }
   const executedMigrations = [
     ...previewMigrations,
     ...knowledgePreviewMigrations,
+    ...savedKnowledgePreviewMigrations,
     ...(yield* run({ loader: makeMigrationLoader(toMigrationInclusive) })),
   ];
   const migrations = executedMigrations.map(([id, name]) => `${id}_${name}`);
diff --git a/apps/server/src/persistence/Migrations/092_ProjectKnowledgeSavedLifecycle.ts b/apps/server/src/persistence/Migrations/092_ProjectKnowledgeSavedLifecycle.ts
new file mode 100644
index 0000000000000..24debeeebed1a
--- /dev/null
+++ b/apps/server/src/persistence/Migrations/092_ProjectKnowledgeSavedLifecycle.ts
@@ -0,0 +1,77 @@
+import * as Effect from "effect/Effect";
+import * as SqlClient from "effect/sql/SqlClient";
+
+export default Effect.gen(function* () {
+  const sql = yield* SqlClient.SqlClient;
+  yield* sql`
+    UPDATE project_knowledge_facts SET supersedes_id = NULL
+    WHERE supersedes_id IS NOT NULL AND NOT EXISTS (
+      SELECT 1 FROM project_knowledge_facts AS target
+      WHERE target.id = project_knowledge_facts.supersedes_id
+        AND target.project_id = project_knowledge_facts.project_id
+        AND target.superseded_by_id = project_knowledge_facts.id
+    )
+  `;
+  yield* sql`
+    CREATE TABLE project_knowledge_facts_saved (
+      id TEXT PRIMARY KEY NOT NULL,
+      project_id TEXT NOT NULL,
+      status TEXT NOT NULL CHECK (status IN ('saved', 'forgotten')),
+      revision INTEGER NOT NULL CHECK (revision > 0),
+      content TEXT,
+      entities_json TEXT NOT NULL,
+      tags_json TEXT NOT NULL,
+      kind TEXT NOT NULL,
+      trust REAL NOT NULL CHECK (trust >= 0 AND trust <= 1),
+      retrieval_count INTEGER NOT NULL DEFAULT 0,
+      created_at TEXT NOT NULL,
+      updated_at TEXT NOT NULL,
+      forgotten_at TEXT,
+      sources_json TEXT NOT NULL,
+      origin_json TEXT NOT NULL,
+      supersedes_id TEXT,
+      superseded_by_id TEXT,
+      CHECK (status != 'forgotten' OR
+        (content IS NULL AND entities_json = '[]' AND tags_json = '[]' AND sources_json = '[]'))
+    )
+  `;
+  yield* sql`
+    INSERT INTO project_knowledge_facts_saved (
+      rowid, id, project_id, status, revision, content, entities_json, tags_json, kind, trust,
+      retrieval_count, created_at, updated_at, forgotten_at, sources_json, origin_json,
+      supersedes_id, superseded_by_id
+    )
+    SELECT rowid, id, project_id,
+      CASE status WHEN 'forgotten' THEN 'forgotten' ELSE 'saved' END,
+      revision, content, entities_json, tags_json, kind, trust,
+      retrieval_count, created_at, updated_at, forgotten_at, sources_json, origin_json,
+      supersedes_id, superseded_by_id
+    FROM project_knowledge_facts
+  `;
+  yield* sql`DROP TABLE project_knowledge_facts`;
+  yield* sql`ALTER TABLE project_knowledge_facts_saved RENAME TO project_knowledge_facts`;
+  yield* sql`CREATE INDEX project_knowledge_facts_project ON project_knowledge_facts(project_id, status, id)`;
+  yield* sql`
+    CREATE TRIGGER project_knowledge_insert AFTER INSERT ON project_knowledge_facts BEGIN
+      INSERT INTO project_knowledge_fts(rowid, content, entities_json, tags_json)
+      VALUES (new.rowid, new.content, new.entities_json, new.tags_json);
+    END
+  `;
+  yield* sql`
+    CREATE TRIGGER project_knowledge_delete AFTER DELETE ON project_knowledge_facts BEGIN
+      INSERT INTO project_knowledge_fts(project_knowledge_fts, rowid, content, entities_json, tags_json)
+      VALUES ('delete', old.rowid, old.content, old.entities_json, old.tags_json);
+    END
+  `;
+  yield* sql`
+    CREATE TRIGGER project_knowledge_update AFTER UPDATE OF content, entities_json, tags_json ON project_knowledge_facts BEGIN
+      INSERT INTO project_knowledge_fts(project_knowledge_fts, rowid, content, entities_json, tags_json)
+      VALUES ('delete', old.rowid, old.content, old.entities_json, old.tags_json);
+      INSERT INTO project_knowledge_fts(rowid, content, entities_json, tags_json)
+      VALUES (new.rowid, new.content, new.entities_json, new.tags_json);
+    END
+  `;
+  yield* sql`INSERT INTO project_knowledge_fts(project_knowledge_fts) VALUES ('rebuild')`;
+  yield* sql`ALTER TABLE project_knowledge_assisted_runs RENAME COLUMN proposed TO saved`;
+  yield* sql`ALTER TABLE project_knowledge_assisted_runs DROP COLUMN suggested`;
+});
diff --git a/apps/server/src/persistence/reconcileSavedKnowledgePreviewMigration.ts b/apps/server/src/persistence/reconcileSavedKnowledgePreviewMigration.ts
new file mode 100644
index 0000000000000..1f8e38553236b
--- /dev/null
+++ b/apps/server/src/persistence/reconcileSavedKnowledgePreviewMigration.ts
@@ -0,0 +1,61 @@
+import * as Effect from "effect/Effect";
+import * as Migrator from "effect/sql/Migrator";
+import * as SqlClient from "effect/sql/SqlClient";
+
+import DropProjectLoomIndexes from "./Migrations/091_DropProjectLoomIndexes.ts";
+
+// The saved-memory preview used 91 before Loom cleanup took that id.
+export const reconcileSavedKnowledgePreviewMigration = Effect.fn(
+  "reconcileSavedKnowledgePreviewMigration",
+)(function* (manifest: ReadonlyArray<readonly [number, string]>) {
+  const sql = yield* SqlClient.SqlClient;
+  return yield* sql.withTransaction(
+    Effect.gen(function* () {
+      const tables = yield* sql`
+        SELECT name FROM sqlite_master WHERE type = 'table' AND name = 'effect_sql_migrations'
+      `;
+      if (tables.length === 0) return [];
+      const history = yield* sql<{
+        readonly migration_id: number;
+        readonly name: string;
+      }>`SELECT migration_id, name FROM effect_sql_migrations WHERE migration_id >= 55`;
+      const recorded = new Map(history.map((row) => [row.migration_id, row.name]));
+      if (recorded.get(91) !== "ProjectKnowledgeSavedLifecycle") return [];
+
+      const expected = new Map(manifest.filter(([id]) => id >= 55 && id < 91));
+      expected.set(91, "ProjectKnowledgeSavedLifecycle");
+      if (
+        [...expected].some(([id, name]) => recorded.get(id) !== name) ||
+        history.some((row) => expected.get(row.migration_id) !== row.name)
+      ) {
+        return yield* new Migrator.MigrationError({
+          kind: "BadState",
+          message: "Cannot reconcile saved-memory preview with unexpected migration history.",
+        });
+      }
+      const facts = yield* sql<{ readonly sql: string }>`
+        SELECT sql FROM sqlite_master WHERE type = 'table' AND name = 'project_knowledge_facts'
+      `;
+      const assistedColumns = yield* sql<{ readonly name: string }>`
+        PRAGMA table_info(project_knowledge_assisted_runs)
+      `;
+      if (
+        !facts[0]?.sql.includes("CHECK (status IN ('saved', 'forgotten'))") ||
+        !assistedColumns.some(({ name }) => name === "saved") ||
+        assistedColumns.some(({ name }) => name === "proposed" || name === "suggested")
+      ) {
+        return yield* new Migrator.MigrationError({
+          kind: "BadState",
+          message: "Cannot reconcile saved-memory preview with an incomplete saved schema.",
+        });
+      }
+
+      yield* DropProjectLoomIndexes;
+      yield* sql`UPDATE effect_sql_migrations SET migration_id = 92 WHERE migration_id = 91`;
+      yield* sql`
+        INSERT INTO effect_sql_migrations (migration_id, name) VALUES (91, 'DropProjectLoomIndexes')
+      `;
+      return [[91, "DropProjectLoomIndexes"]] as const;
+    }),
+  );
+});
diff --git a/apps/server/src/persistence/reconcileV2PreviewMigration.test.ts b/apps/server/src/persistence/reconcileV2PreviewMigration.test.ts
index 377a081228247..5ce01dedc80cd 100644
--- a/apps/server/src/persistence/reconcileV2PreviewMigration.test.ts
+++ b/apps/server/src/persistence/reconcileV2PreviewMigration.test.ts
@@ -71,6 +71,7 @@ describe("V2 preview upgrade", () => {
         [89, "DelegatedRecoveryIndexes"],
         [90, "ScheduledTaskThreadMode"],
         [91, "DropProjectLoomIndexes"],
+        [92, "ProjectKnowledgeSavedLifecycle"],
       ]);
       assert.deepStrictEqual(yield* runMigrations(), []);
       assert.deepStrictEqual(yield* sql`SELECT * FROM orchestration_v2_legacy_imports`, imports);
@@ -185,6 +186,7 @@ describe("V2 preview upgrade", () => {
         [89, "DelegatedRecoveryIndexes"],
         [90, "ScheduledTaskThreadMode"],
         [91, "DropProjectLoomIndexes"],
+        [92, "ProjectKnowledgeSavedLifecycle"],
       ]);
     }).pipe(Effect.provide(NodeSqliteClient.layer({ filename: ":memory:" }))),
   );
diff --git a/apps/server/src/project/knowledge/AssistedLearning.test.ts b/apps/server/src/project/knowledge/AssistedLearning.test.ts
index 6cd406a681452..cc231cce7aa75 100644
--- a/apps/server/src/project/knowledge/AssistedLearning.test.ts
+++ b/apps/server/src/project/knowledge/AssistedLearning.test.ts
@@ -45,9 +45,11 @@ const threadId = ThreadId.make("learning-thread");
 const runId = RunId.make("learning-run");
 const providerInstanceId = ProviderInstanceId.make("codex");
 const at = DateTime.makeUnsafe(Date.UTC(2026, 9, 1));
-const optedIn = { projectSettingsOverrides: { [projectId]: { enableAssistedLearning: true } } };
 
-const makeLayer = (generate: Generate, settings: Parameters<typeof ServerSettings.layerTest>[0]) =>
+const makeLayer = (
+  generate: Generate,
+  settings: Parameters<typeof ServerSettings.layerTest>[0] = {},
+) =>
   AssistedLearning.layerRunner.pipe(
     Layer.provide(
       Layer.mock(TextGeneration.TextGeneration)({ generateKnowledgeCandidates: generate }),
@@ -208,15 +210,26 @@ const snakeCase = {
   citations: ["M1", "M2"],
 };
 
-it("schedules one coalesced job per run only for opted-in projects", () => {
+it("schedules one coalesced job per run unless the environment or project opted out", () => {
   const input = { projectId, threadId, runId, status: "completed" as const };
-  assert.deepEqual(assistedLearningEffects({ ...input, settings: DEFAULT_SERVER_SETTINGS }), []);
-  const otherProject = {
-    ...DEFAULT_SERVER_SETTINGS,
-    projectSettingsOverrides: { [ProjectId.make("other")]: { enableAssistedLearning: true } },
-  };
-  assert.deepEqual(assistedLearningEffects({ ...input, settings: otherProject }), []);
-  const settings = { ...DEFAULT_SERVER_SETTINGS, ...optedIn };
+  const settings = DEFAULT_SERVER_SETTINGS;
+  assert.deepEqual(
+    assistedLearningEffects({
+      ...input,
+      settings: { ...DEFAULT_SERVER_SETTINGS, enableAssistedLearning: false },
+    }),
+    [],
+  );
+  assert.deepEqual(
+    assistedLearningEffects({
+      ...input,
+      settings: {
+        ...DEFAULT_SERVER_SETTINGS,
+        projectSettingsOverrides: { [projectId]: { enableAssistedLearning: false } },
+      },
+    }),
+    [],
+  );
   const [effect] = assistedLearningEffects({ ...input, settings });
   assert.equal(effect?.id, `effect:project-knowledge.extract:${runId}`);
   assert.deepEqual(effect?.request, {
@@ -230,7 +243,7 @@ it("schedules one coalesced job per run only for opted-in projects", () => {
   );
 });
 
-it.effect("turns a settled run into cited proposals exactly once", () => {
+it.effect("saves a settled run's cited facts exactly once, recallable without review", () => {
   const model = scripted([
     { candidates: [snakeCase, { ...snakeCase, content: "Invented.", citations: ["M9"] }] },
   ]);
@@ -249,10 +262,10 @@ it.effect("turns a settled run into cited proposals exactly once", () => {
         ["M2", "assistant"],
       ],
     );
-    const { facts } = yield* knowledge.list({ projectId, status: "proposed", origin: "assisted" });
+    const { facts } = yield* knowledge.list({ projectId, status: "saved", origin: "assisted" });
     assert.equal(facts.length, 1);
     const fact = facts[0]!;
-    assert.equal(fact.status, "proposed");
+    assert.equal(fact.status, "saved");
     assert.equal(fact.content, snakeCase.content);
     assert.deepEqual(fact.origin, { kind: "assisted", threadId, runId, runOutcome: "completed" });
     assert.deepEqual(fact.sources, [
@@ -270,20 +283,18 @@ it.effect("turns a settled run into cited proposals exactly once", () => {
       },
     ]);
     const status = yield* knowledge.status({ projectId });
-    assert.equal(status.assistedLearning?.suggested, 1);
     assert.deepEqual(
       status.assistedLearning?.lastRun && {
         result: status.assistedLearning.lastRun.result,
-        proposed: status.assistedLearning.lastRun.proposed,
+        saved: status.assistedLearning.lastRun.saved,
         reason: status.assistedLearning.lastRun.reason,
       },
       {
         result: "ran",
-        proposed: 1,
+        saved: 1,
         reason: "1 candidate dropped: citations did not verify or exceeded the limit.",
       },
     );
-    // Unreviewed model text stays out of agent recall.
     const recalled = yield* knowledge.search({
       projectId,
       query: "snake_case API responses",
@@ -291,9 +302,14 @@ it.effect("turns a settled run into cited proposals exactly once", () => {
     });
     assert.deepEqual(
       recalled.facts.state === "available" ? recalled.facts.matches.map((m) => m.fact.id) : [],
-      [],
+      [fact.id],
     );
-  }).pipe(Effect.provide(makeLayer(model.generate, optedIn)));
+    const { history } = yield* knowledge.readFact({ projectId, factId: fact.id });
+    assert.deepEqual(
+      history.map((entry) => entry.event),
+      ["created"],
+    );
+  }).pipe(Effect.provide(makeLayer(model.generate)));
 });
 
 it.effect("marks candidates from a failed run with that outcome", () => {
@@ -304,81 +320,141 @@ it.effect("marks candidates from a failed run with that outcome", () => {
     const knowledge = yield* ProjectKnowledge.ProjectKnowledge;
     yield* runner.extractRun({ threadId, runId, runOutcome: "failed" });
     assert.equal(model.calls[0]!.runOutcome, "failed");
-    const { facts } = yield* knowledge.list({ projectId, status: "proposed" });
+    const { facts } = yield* knowledge.list({ projectId, status: "saved" });
     assert.deepEqual(
       facts.map((fact) => [fact.status, fact.origin.kind === "assisted" && fact.origin.runOutcome]),
-      [["proposed", "failed"]],
+      [["saved", "failed"]],
     );
-  }).pipe(Effect.provide(makeLayer(model.generate, optedIn)));
+  }).pipe(Effect.provide(makeLayer(model.generate)));
 });
 
-it.effect("suggests replacing a near duplicate instead of duplicating it", () => {
-  const model = scripted([
-    {
-      candidates: [
-        { ...snakeCase, content: "API responses use snake_case keys for every public endpoint." },
-        { ...snakeCase, content: "Tests run with vp test." },
-      ],
-    },
-  ]);
-  return Effect.gen(function* () {
-    yield* seedRun;
-    const runner = yield* AssistedLearningRunner;
-    const knowledge = yield* ProjectKnowledge.ProjectKnowledge;
-    const existing = yield* knowledge.mutate(
+it.effect(
+  "skips exact duplicates and saves different facts without replacing existing ones",
+  () => {
+    const conflict = "The primary database store switched to Postgres instead of MySQL.";
+    const model = scripted([
       {
-        action: "create",
-        commandId: CommandId.make("existing"),
-        projectId,
-        factId: ProjectFactId.make("existing"),
-        content: "API responses use snake_case keys for every endpoint.",
-        entities: [],
-        tags: [],
-        kind: "convention",
-        sources: [{ kind: "user" }],
+        candidates: [
+          { ...snakeCase, content: "API responses use snake_case keys for every public endpoint." },
+          { ...snakeCase, content: "Tests run with vp test." },
+          { ...snakeCase, content: conflict },
+        ],
       },
-      { kind: "user" },
-    );
-    yield* knowledge.mutate(
-      {
-        action: "create",
-        commandId: CommandId.make("exact"),
+    ]);
+    return Effect.gen(function* () {
+      yield* seedRun;
+      const runner = yield* AssistedLearningRunner;
+      const knowledge = yield* ProjectKnowledge.ProjectKnowledge;
+      const create = (id: string, content: string) => ({
+        action: "create" as const,
+        commandId: CommandId.make(id),
         projectId,
-        factId: ProjectFactId.make("exact"),
-        content: "Tests run with vp test.",
+        factId: ProjectFactId.make(id),
+        content,
         entities: [],
         tags: [],
-        kind: "convention",
-        sources: [{ kind: "user" }],
-      },
-      { kind: "user" },
-    );
-    yield* runner.extractRun({ threadId, runId, runOutcome: "completed" });
-    const { facts } = yield* knowledge.list({ projectId, status: "proposed", origin: "assisted" });
-    // The exact duplicate created nothing; the near duplicate names what it would replace.
-    assert.deepEqual(
-      facts.map((fact) => [fact.content, fact.supersedes]),
-      [["API responses use snake_case keys for every public endpoint.", existing.id]],
-    );
-    assert.equal((yield* knowledge.status({ projectId })).assistedLearning?.lastRun?.suggested, 1);
-    // The existing fact stays recalled until a user accepts the replacement.
-    const before = (yield* knowledge.readFact({ projectId, factId: existing.id })).fact;
-    assert.equal(before.supersededBy ?? null, null);
-    const suggestion = facts[0]!;
-    yield* knowledge.mutate(
-      {
-        action: "confirm",
-        commandId: CommandId.make("accept"),
+        kind: "convention" as const,
+        sources: [{ kind: "user" as const }],
+      });
+      const seed = (id: string, content: string) =>
+        knowledge.mutate(create(id, content), { kind: "user" });
+      const nearDuplicate = yield* seed(
+        "near",
+        "API responses use snake_case keys for every endpoint.",
+      );
+      yield* seed("exact", "Tests run with vp test.");
+      const database = yield* seed("database", "The primary database store is MySQL or Postgres.");
+      yield* runner.extractRun({ threadId, runId, runOutcome: "completed" });
+      const { facts } = yield* knowledge.list({ projectId, status: "saved", origin: "assisted" });
+      assert.deepEqual(
+        facts.map((fact) => [fact.content, fact.supersedes ?? null]).toSorted(),
+        [
+          ["API responses use snake_case keys for every public endpoint.", null],
+          [conflict, null],
+        ].toSorted(),
+      );
+      const { lastRun } = (yield* knowledge.status({ projectId })).assistedLearning!;
+      assert.deepEqual([lastRun?.saved, lastRun?.reason], [2, "1 duplicate skipped."]);
+      for (const existing of [nearDuplicate, database]) {
+        const { fact } = yield* knowledge.readFact({ projectId, factId: existing.id });
+        assert.equal(fact.supersededBy ?? null, null);
+        assert.equal(fact.revision, 1);
+      }
+      // Exact matching cannot depend on the three UI hints or the forty lexical neighbours.
+      for (let index = 0; index < 101; index++) {
+        yield* seed(
+          `crowded-${String(index).padStart(3, "0")}`,
+          `workers restart then services deploy${"!".repeat(index + 1)}`,
+        );
+      }
+      yield* seed("z-crowded-exact", "deploy services then restart workers");
+      yield* seed("unicode-exact", "测试需要先运行单元测试。");
+      for (const [index, content] of [
+        "  DEPLOY   services then RESTART workers\n",
+        "测试需要先运行单元测试。",
+      ].entries()) {
+        const result = yield* knowledge.saveAutomatic(
+          { ...create(`crowded-auto-${index}`, content), sources: facts[0]!.sources },
+          { kind: "assisted", threadId, runId, runOutcome: "completed" },
+        );
+        assert.equal(result.outcome, "duplicate");
+      }
+      const recalled = yield* knowledge.search({
         projectId,
-        factId: suggestion.id,
-        expectedRevision: suggestion.revision,
-      },
-      { kind: "user" },
-    );
-    const after = (yield* knowledge.readFact({ projectId, factId: existing.id })).fact;
-    assert.equal(after.supersededBy, suggestion.id);
-  }).pipe(Effect.provide(makeLayer(model.generate, optedIn)));
-});
+        query: "primary database store",
+        requestId: CommandId.make("recall-conflict"),
+      });
+      assert.deepEqual(
+        recalled.facts.state === "available"
+          ? recalled.facts.matches.map((match) => match.fact.id).toSorted()
+          : [],
+        [database.id, facts.find((fact) => fact.content === conflict)!.id].toSorted(),
+      );
+
+      const positive = yield* seed(
+        "positive",
+        "API responses use snake_case keys for every public endpoint.",
+      );
+      yield* seed("retries", "The worker pool never retries failed jobs.");
+      yield* seed("approval", "Deploys require a manual approval step.");
+      yield* seed("blocked", "The worker is not idle but is blocked.");
+      yield* seed("enabled", "The worker pool is enabled for every public endpoint.");
+      const negated = "API responses do not use snake_case keys for every public endpoint.";
+      const outcomes = yield* Effect.forEach(
+        [
+          negated,
+          negated,
+          "API responses do not use snake_case keys for every single public endpoint.",
+          "The worker pool retries failed jobs.",
+          "Deploys don't require a manual approval step.",
+          "The worker is idle but is not blocked.",
+          "The worker pool is disabled for every public endpoint.",
+        ],
+        (content, index) =>
+          knowledge
+            .saveAutomatic(
+              { ...create(`negation-${index}`, content), sources: facts[0]!.sources },
+              { kind: "assisted", threadId, runId, runOutcome: "completed" },
+            )
+            .pipe(Effect.map((result) => result.outcome)),
+      );
+      assert.deepEqual(outcomes, [
+        "saved",
+        "duplicate",
+        "saved",
+        "saved",
+        "saved",
+        "saved",
+        "saved",
+      ]);
+      const { fact: kept } = yield* knowledge.readFact({ projectId, factId: positive.id });
+      assert.deepEqual(
+        [kept.revision, kept.supersededBy ?? null, kept.supersedes ?? null],
+        [1, null, null],
+      );
+    }).pipe(Effect.provide(makeLayer(model.generate)));
+  },
+);
 
 const testsCommand = {
   kind: "convention" as const,
@@ -387,7 +463,7 @@ const testsCommand = {
   citations: ["M2"],
 };
 
-it.effect("does not bring back a forgotten proposal when the job runs again", () => {
+it.effect("does not bring back a forgotten fact when the job runs again", () => {
   // The retried model call returns the same candidate at a different position.
   const model = scripted([{ candidates: [snakeCase] }, { candidates: [testsCommand, snakeCase] }]);
   return Effect.gen(function* () {
@@ -396,14 +472,14 @@ it.effect("does not bring back a forgotten proposal when the job runs again", ()
     const knowledge = yield* ProjectKnowledge.ProjectKnowledge;
     const sql = yield* SqlClient.SqlClient;
     yield* runner.extractRun({ threadId, runId, runOutcome: "completed" });
-    const [proposal] = (yield* knowledge.list({ projectId, status: "proposed" })).facts;
+    const [saved] = (yield* knowledge.list({ projectId, status: "saved" })).facts;
     yield* knowledge.mutate(
       {
         action: "forget",
-        commandId: CommandId.make("forget-proposal"),
+        commandId: CommandId.make("forget-saved"),
         projectId,
-        factId: proposal!.id,
-        expectedRevision: proposal!.revision,
+        factId: saved!.id,
+        expectedRevision: saved!.revision,
       },
       { kind: "user" },
     );
@@ -412,24 +488,24 @@ it.effect("does not bring back a forgotten proposal when the job runs again", ()
     yield* runner.extractRun({ threadId, runId, runOutcome: "completed" });
     assert.equal(model.calls.length, 2);
     assert.deepEqual(
-      (yield* knowledge.list({ projectId, status: "proposed" })).facts.map((fact) => fact.content),
+      (yield* knowledge.list({ projectId, status: "saved" })).facts.map((fact) => fact.content),
       [testsCommand.content],
     );
     assert.deepEqual(
       (yield* knowledge.list({ projectId, status: "forgotten" })).facts.map((fact) => fact.id),
-      [proposal!.id],
+      [saved!.id],
     );
-  }).pipe(Effect.provide(makeLayer(model.generate, optedIn)));
+  }).pipe(Effect.provide(makeLayer(model.generate)));
 });
 
-it.effect("commits a run's proposals only together with its outcome", () => {
+it.effect("commits a run's facts only together with its outcome", () => {
   const model = scripted([{ candidates: [snakeCase, testsCommand] }]);
   return Effect.gen(function* () {
     yield* seedRun;
     const runner = yield* AssistedLearningRunner;
     const knowledge = yield* ProjectKnowledge.ProjectKnowledge;
     const sql = yield* SqlClient.SqlClient;
-    const proposals = Effect.map(knowledge.list({ projectId, status: "proposed" }), (listed) =>
+    const savedContent = Effect.map(knowledge.list({ projectId, status: "saved" }), (listed) =>
       listed.facts.map((fact) => fact.content).toSorted(),
     );
     yield* sql`CREATE TEMP TRIGGER lose_outcome BEFORE INSERT ON project_knowledge_assisted_runs
@@ -438,16 +514,16 @@ it.effect("commits a run's proposals only together with its outcome", () => {
       .extractRun({ threadId, runId, runOutcome: "completed" })
       .pipe(Effect.result);
     assert.isTrue(Result.isFailure(lost));
-    // Nothing a user could review, and so forget, exists without the outcome.
-    assert.deepEqual(yield* proposals, []);
+    // Nothing a user could edit or forget exists without the outcome.
+    assert.deepEqual(yield* savedContent, []);
     yield* sql`DROP TRIGGER lose_outcome`;
     yield* runner.extractRun({ threadId, runId, runOutcome: "completed" });
-    assert.deepEqual(yield* proposals, [snakeCase.content, testsCommand.content].toSorted());
-    assert.equal((yield* knowledge.status({ projectId })).assistedLearning?.lastRun?.proposed, 2);
-  }).pipe(Effect.provide(makeLayer(model.generate, optedIn)));
+    assert.deepEqual(yield* savedContent, [snakeCase.content, testsCommand.content].toSorted());
+    assert.equal((yield* knowledge.status({ projectId })).assistedLearning?.lastRun?.saved, 2);
+  }).pipe(Effect.provide(makeLayer(model.generate)));
 });
 
-it.effect("proposes nothing for a run another attempt finished while the model ran", () => {
+it.effect("saves nothing for a run another attempt finished while the model ran", () => {
   let finishElsewhere: Effect.Effect<void> = Effect.void;
   const generate: Generate = () =>
     Effect.suspend(() => finishElsewhere).pipe(Effect.as({ candidates: [snakeCase] }));
@@ -462,14 +538,13 @@ it.effect("proposes nothing for a run another attempt finished while the model r
         runOutcome: "completed",
         result: "ran",
         reason: null,
-        proposed: 0,
-        suggested: 0,
+        saved: 0,
         at: DateTime.formatIso(at),
       })
       .pipe(Effect.orDie);
     yield* runner.extractRun({ threadId, runId, runOutcome: "completed" });
-    assert.deepEqual((yield* knowledge.list({ projectId, status: "proposed" })).facts, []);
-  }).pipe(Effect.provide(makeLayer(generate, optedIn)));
+    assert.deepEqual((yield* knowledge.list({ projectId, status: "saved" })).facts, []);
+  }).pipe(Effect.provide(makeLayer(generate)));
 });
 
 it.effect("drops a candidate citing more messages than a fact may carry", () => {
@@ -488,18 +563,18 @@ it.effect("drops a candidate citing more messages than a fact may carry", () =>
     const knowledge = yield* ProjectKnowledge.ProjectKnowledge;
     yield* runner.extractRun({ threadId, runId, runOutcome: "completed" });
     assert.deepEqual(
-      (yield* knowledge.list({ projectId, status: "proposed" })).facts.map((fact) => fact.content),
+      (yield* knowledge.list({ projectId, status: "saved" })).facts.map((fact) => fact.content),
       [testsCommand.content],
     );
     const lastRun = (yield* knowledge.status({ projectId })).assistedLearning?.lastRun;
     assert.deepEqual(
-      [lastRun?.result, lastRun?.proposed, lastRun?.reason],
+      [lastRun?.result, lastRun?.saved, lastRun?.reason],
       ["ran", 1, "1 candidate dropped: citations did not verify or exceeded the limit."],
     );
-  }).pipe(Effect.provide(makeLayer(model.generate, optedIn)));
+  }).pipe(Effect.provide(makeLayer(model.generate)));
 });
 
-it.effect("records skipped and failed outcomes without proposing anything", () => {
+it.effect("records skipped and failed outcomes without saving anything", () => {
   const failing = scripted([
     new TextGenerationError({ operation: "generateKnowledgeCandidates", detail: "Rate limited." }),
   ]);
@@ -510,11 +585,11 @@ it.effect("records skipped and failed outcomes without proposing anything", () =
     yield* runner.extractRun({ threadId, runId, runOutcome: "completed" });
     const lastRun = (yield* knowledge.status({ projectId })).assistedLearning?.lastRun;
     assert.deepEqual([lastRun?.result, lastRun?.reason], ["failed", "Rate limited."]);
-    assert.deepEqual((yield* knowledge.list({ projectId, status: "proposed" })).facts, []);
-  }).pipe(Effect.provide(makeLayer(failing.generate, optedIn)));
+    assert.deepEqual((yield* knowledge.list({ projectId, status: "saved" })).facts, []);
+  }).pipe(Effect.provide(makeLayer(failing.generate)));
 });
 
-it.effect("skips a run whose project turned assisted learning off before it ran", () => {
+it.effect("skips a run whose project turned automatic memory off before it ran", () => {
   const model = scripted([{ candidates: [snakeCase] }]);
   return Effect.gen(function* () {
     yield* seedRun;
@@ -526,5 +601,11 @@ it.effect("skips a run whose project turned assisted learning off before it ran"
       (yield* knowledge.status({ projectId })).assistedLearning?.lastRun?.result,
       "skipped",
     );
-  }).pipe(Effect.provide(makeLayer(model.generate, {})));
+  }).pipe(
+    Effect.provide(
+      makeLayer(model.generate, {
+        projectSettingsOverrides: { [projectId]: { enableAssistedLearning: false } },
+      }),
+    ),
+  );
 });
diff --git a/apps/server/src/project/knowledge/AssistedLearning.ts b/apps/server/src/project/knowledge/AssistedLearning.ts
index 9175ec85c8096..c224f63d74ed6 100644
--- a/apps/server/src/project/knowledge/AssistedLearning.ts
+++ b/apps/server/src/project/knowledge/AssistedLearning.ts
@@ -27,7 +27,6 @@ import {
 } from "./assistedLearningSchedule.ts";
 import * as ProjectKnowledge from "./ProjectKnowledge.ts";
 
-// Bounds on what one run sends to the model and what it may propose back.
 const MAX_MESSAGES = 24;
 const MAX_MESSAGE_CHARS = 6_000;
 const MAX_TRANSCRIPT_CHARS = 32_000;
@@ -35,8 +34,6 @@ const MAX_CANDIDATES = 3;
 const MAX_CANDIDATE_CHARS = 2_000;
 const MAX_ENTITIES = 5;
 
-// The provider schema keeps the reply loosely typed; this is the boundary that decides what
-// may become a proposal.
 const Candidate = Schema.Struct({
   kind: Schema.Literals(["decision", "constraint", "convention"]),
   content: Schema.Trim.check(Schema.isNonEmpty(), Schema.isMaxLength(MAX_CANDIDATE_CHARS)),
@@ -44,7 +41,6 @@ const Candidate = Schema.Struct({
   citations: Schema.Array(Schema.String),
 });
 const decodeCandidate = Schema.decodeUnknownOption(Candidate);
-// The proposal must also satisfy the mutation contract, which caps citations per fact.
 const isMutation = Schema.is(ProjectKnowledgeMutation);
 
 const make = Effect.gen(function* () {
@@ -70,14 +66,7 @@ const make = Effect.gen(function* () {
       if (records === null) return;
       const projectId: ProjectId = records.thread.projectId;
       if (yield* knowledge.hasAssistedRun(projectId, runId)) return;
-      const finish = (
-        result: AssistedLearningRun["result"],
-        reason: string | null,
-        counts: { readonly proposed: number; readonly suggested: number } = {
-          proposed: 0,
-          suggested: 0,
-        },
-      ) =>
+      const finish = (result: AssistedLearningRun["result"], reason: string | null, saved = 0) =>
         DateTime.now.pipe(
           Effect.flatMap((now) =>
             knowledge.recordAssistedRun(projectId, {
@@ -86,7 +75,7 @@ const make = Effect.gen(function* () {
               runOutcome,
               result,
               reason,
-              ...counts,
+              saved,
               at: DateTime.formatIso(now),
             }),
           ),
@@ -98,7 +87,7 @@ const make = Effect.gen(function* () {
       ).settings;
       // The setting can change between scheduling and running; the latest choice wins.
       if (!settings.enableAssistedLearning)
-        return yield* finish("skipped", "Assisted learning is off for this project.");
+        return yield* finish("skipped", "Automatic memory is off for this project.");
       const finalized = records.messages
         .filter(
           (message) =>
@@ -164,13 +153,11 @@ const make = Effect.gen(function* () {
         );
       if (!generated.ok) return yield* finish("failed", generated.detail.slice(0, 300));
       const actor = { kind: "assisted" as const, threadId, runId, runOutcome };
-      // Proposals commit with the run's outcome. A lost outcome then leaves nothing a user could
-      // have reviewed or forgotten, and an attempt that finished first keeps its result.
       yield* sql.withTransaction(
         Effect.gen(function* () {
           if (yield* knowledge.hasAssistedRun(projectId, runId)) return;
-          let proposed = 0;
-          let suggested = 0;
+          let saved = 0;
+          let duplicates = 0;
           let dropped = 0;
           for (const raw of generated.candidates.slice(0, MAX_CANDIDATES)) {
             const candidate = Option.getOrNull(decodeCandidate(raw));
@@ -184,13 +171,11 @@ const make = Effect.gen(function* () {
               dropped += 1;
               continue;
             }
-            // Named by run and content, so a retry that regenerates candidates in another order
-            // lands on the same ids and cannot bring back a proposal that was since forgotten.
-            const id = `assisted:${runId}:${(yield* digest(candidate.content)).slice(0, 32)}`;
-            const proposal = {
+            const runContentId = `assisted:${runId}:${(yield* digest(candidate.content)).slice(0, 32)}`;
+            const create = {
               action: "create" as const,
-              commandId: CommandId.make(id),
-              factId: ProjectFactId.make(id),
+              commandId: CommandId.make(runContentId),
+              factId: ProjectFactId.make(runContentId),
               projectId,
               kind: candidate.kind,
               content: candidate.content,
@@ -204,11 +189,11 @@ const make = Effect.gen(function* () {
               tags: [],
               sources,
             };
-            if (!isMutation(proposal)) {
+            if (!isMutation(create)) {
               dropped += 1;
               continue;
             }
-            const result = yield* knowledge.proposeAssisted(proposal, actor).pipe(
+            const result = yield* knowledge.saveAutomatic(create, actor).pipe(
               Effect.catchTags({
                 ProjectKnowledgeError: (error) =>
                   error.reason === "storage" || error.reason === "missing_project"
@@ -216,20 +201,24 @@ const make = Effect.gen(function* () {
                     : Effect.succeed({ outcome: "dropped" as const }),
               }),
             );
-            if (result.outcome === "proposed") proposed += 1;
-            else if (result.outcome === "suggested") {
-              proposed += 1;
-              suggested += 1;
-            } else if (result.outcome === "dropped") dropped += 1;
+            if (result.outcome === "saved") saved += 1;
+            else if (result.outcome === "duplicate") duplicates += 1;
+            else if (result.outcome === "dropped") dropped += 1;
           }
+          const notes = [
+            ...(duplicates > 0
+              ? [`${duplicates} duplicate${duplicates === 1 ? "" : "s"} skipped.`]
+              : []),
+            ...(dropped > 0
+              ? [
+                  `${dropped} candidate${dropped === 1 ? "" : "s"} dropped: citations did not verify or exceeded the limit.`,
+                ]
+              : []),
+          ];
           yield* finish(
             "ran",
-            proposed === 0 && dropped === 0
-              ? "Nothing durable to propose."
-              : dropped > 0
-                ? `${dropped} candidate${dropped === 1 ? "" : "s"} dropped: citations did not verify or exceeded the limit.`
-                : null,
-            { proposed, suggested },
+            notes.length > 0 ? notes.join(" ") : saved === 0 ? "Nothing durable to save." : null,
+            saved,
           );
         }),
       );
diff --git a/apps/server/src/project/knowledge/ProjectKnowledge.test.ts b/apps/server/src/project/knowledge/ProjectKnowledge.test.ts
index 1c58dee565ec2..a709b2fafeb6d 100644
--- a/apps/server/src/project/knowledge/ProjectKnowledge.test.ts
+++ b/apps/server/src/project/knowledge/ProjectKnowledge.test.ts
@@ -237,7 +237,7 @@ it.effect("renders user attribution once in prepared and provider fact headers",
       const header = text.match(/\[Fact [^\n]+\]/)?.[0];
       assert.isDefined(header);
       assert.equal(header?.match(/User assertion/g)?.length, 1);
-      assert.include(header!, "revision:1 status:confirmed trust:0.5; User assertion]");
+      assert.include(header!, "revision:1 status:saved trust:0.5; User assertion]");
     }
   }).pipe(Effect.provide(layerTest)),
 );
@@ -251,7 +251,7 @@ it.effect("persists edits with CAS, erases forgotten content and cannot resurrec
     const command = createFact(projectId);
     const saved = yield* knowledge.mutate(command, { kind: "user" });
     assert.equal(saved.revision, 1);
-    assert.equal(saved.status, "confirmed");
+    assert.equal(saved.status, "saved");
     const oversized = yield* knowledge
       .mutate(
         { ...createFact(projectId, "oversized"), content: "測".repeat(6000) },
@@ -322,96 +322,78 @@ it.effect("persists edits with CAS, erases forgotten content and cannot resurrec
   }).pipe(Effect.provide(layerTest)),
 );
 
-it.effect(
-  "recalls agent proposals before optional human confirmation and rehydrates exact sources",
-  () =>
-    Effect.gen(function* () {
-      const knowledge = yield* ProjectKnowledge.ProjectKnowledge;
-      const projections = yield* ProjectionStore.ProjectionStoreV2;
-      const projectId = ProjectId.make("sourced");
-      const threadId = ThreadId.make("source-thread");
-      const messageId = MessageId.make("source-message");
-      yield* createProject(projectId);
-      yield* projections.apply(thread(threadId, projectId));
-      yield* projections.apply(
-        message(threadId, messageId, "user", "Use tenant IDs in every cache key."),
-      );
-      const recalled = yield* knowledge.search({
-        projectId,
-        query: "tenant IDs",
-        requestId: CommandId.make("recall-source"),
-      });
-      assert.equal(recalled.messages.matches[0]?.messageId, messageId);
-      const source = yield* knowledge.readSource({
-        projectId,
-        threadId,
-        messageId,
-        ...(recalled.messages.matches[0]?.digest === undefined
-          ? {}
-          : { digest: recalled.messages.matches[0].digest }),
-      });
-      assert.equal(source.state, "available");
-      if (source.state !== "available") return;
-      const proposed = yield* knowledge.mutate(
-        { ...createFact(projectId), sources: [source.source] },
-        { kind: "agent-client", sessionId: "test-agent", label: "Test agent" },
-      );
-      assert.equal(proposed.status, "proposed");
-      assert.deepEqual((yield* knowledge.list({ projectId })).facts, []);
-      const context = {
-        environmentId,
-        projectId,
-        selections: [{ kind: "fact" as const, factId: proposed.id, revision: 1 }],
-        maxBytes: 4096,
-      };
-      const unconfirmed = yield* knowledge.prepareContext(context);
-      assert.include(unconfirmed.text, "status:proposed");
-      const confirm = {
-        action: "confirm" as const,
-        commandId: CommandId.make("confirm"),
-        projectId,
-        factId: proposed.id,
-        expectedRevision: 1,
-      };
-      const forbidden = yield* knowledge
-        .mutate(confirm, { kind: "agent-client", sessionId: "test-agent", label: "Test agent" })
-        .pipe(Effect.result);
-      assert.equal(
-        Result.isFailure(forbidden) && forbidden.failure.reason,
-        "user_confirmation_required",
-      );
-      const confirmed = yield* knowledge.mutate(confirm, { kind: "user" });
-      const preview = yield* knowledge.prepareContext({
-        ...context,
-        selections: [{ kind: "fact", factId: confirmed.id, revision: confirmed.revision }],
-      });
-      assert.include(preview.text, "Cache keys include tenant IDs.");
-      assert.include(preview.text, `Message ${threadId}/${messageId}`);
-      assert.isAtLeast(preview.byteLength, Buffer.byteLength(preview.text));
-      assert.include(
-        (yield* knowledge.resolveForTurn(preview.context)).text,
-        "Cache keys include tenant IDs.",
-      );
-      yield* projections.apply(
-        message(threadId, messageId, "user", "Use tenant IDs and regions.", { minute: 4 }),
-      );
-      const staleSource = yield* knowledge.readSource({ projectId, ...source.source });
-      assert.deepEqual(staleSource, { state: "unavailable", reason: "changed" });
-      const stale = yield* knowledge.resolveForTurn(preview.context).pipe(Effect.result);
-      assert.equal(Result.isFailure(stale) && stale.failure.reason, "stale_context");
-      const records = yield* projections.getThreadRecords(threadId, []);
-      yield* projections.apply({
-        id: EventId.make("delete-source"),
-        threadId,
-        type: "thread.deleted",
-        occurredAt: at(5),
-        payload: { ...records.thread, deletedAt: at(5) },
-      });
-      assert.deepEqual(yield* knowledge.readSource({ projectId, threadId, messageId }), {
-        state: "unavailable",
-        reason: "deleted",
-      });
-    }).pipe(Effect.provide(layerTest)),
+it.effect("saves agent facts for immediate recall and rehydrates exact sources", () =>
+  Effect.gen(function* () {
+    const knowledge = yield* ProjectKnowledge.ProjectKnowledge;
+    const projections = yield* ProjectionStore.ProjectionStoreV2;
+    const projectId = ProjectId.make("sourced");
+    const threadId = ThreadId.make("source-thread");
+    const messageId = MessageId.make("source-message");
+    yield* createProject(projectId);
+    yield* projections.apply(thread(threadId, projectId));
+    yield* projections.apply(
+      message(threadId, messageId, "user", "Use tenant IDs in every cache key."),
+    );
+    const recalled = yield* knowledge.search({
+      projectId,
+      query: "tenant IDs",
+      requestId: CommandId.make("recall-source"),
+    });
+    assert.equal(recalled.messages.matches[0]?.messageId, messageId);
+    const source = yield* knowledge.readSource({
+      projectId,
+      threadId,
+      messageId,
+      ...(recalled.messages.matches[0]?.digest === undefined
+        ? {}
+        : { digest: recalled.messages.matches[0].digest }),
+    });
+    assert.equal(source.state, "available");
+    if (source.state !== "available") return;
+    const saved = yield* knowledge.mutate(
+      { ...createFact(projectId), sources: [source.source] },
+      { kind: "agent-client", sessionId: "test-agent", label: "Test agent" },
+    );
+    assert.equal(saved.status, "saved");
+    assert.deepEqual(
+      (yield* knowledge.list({ projectId })).facts.map((fact) => fact.id),
+      [saved.id],
+    );
+    const context = {
+      environmentId,
+      projectId,
+      selections: [{ kind: "fact" as const, factId: saved.id, revision: 1 }],
+      maxBytes: 4096,
+    };
+    const preview = yield* knowledge.prepareContext(context);
+    assert.include(preview.text, "status:saved");
+    assert.include(preview.text, "Cache keys include tenant IDs.");
+    assert.include(preview.text, `Message ${threadId}/${messageId}`);
+    assert.isAtLeast(preview.byteLength, Buffer.byteLength(preview.text));
+    assert.include(
+      (yield* knowledge.resolveForTurn(preview.context)).text,
+      "Cache keys include tenant IDs.",
+    );
+    yield* projections.apply(
+      message(threadId, messageId, "user", "Use tenant IDs and regions.", { minute: 4 }),
+    );
+    const staleSource = yield* knowledge.readSource({ projectId, ...source.source });
+    assert.deepEqual(staleSource, { state: "unavailable", reason: "changed" });
+    const stale = yield* knowledge.resolveForTurn(preview.context).pipe(Effect.result);
+    assert.equal(Result.isFailure(stale) && stale.failure.reason, "stale_context");
+    const records = yield* projections.getThreadRecords(threadId, []);
+    yield* projections.apply({
+      id: EventId.make("delete-source"),
+      threadId,
+      type: "thread.deleted",
+      occurredAt: at(5),
+      payload: { ...records.thread, deletedAt: at(5) },
+    });
+    assert.deepEqual(yield* knowledge.readSource({ projectId, threadId, messageId }), {
+      state: "unavailable",
+      reason: "deleted",
+    });
+  }).pipe(Effect.provide(layerTest)),
 );
 
 it.effect("lists which facts cite sources that were edited, deleted or archived", () =>
@@ -573,7 +555,7 @@ it.effect(
     }).pipe(Effect.provide(layerTest)),
 );
 
-it.effect("MCP rejects unauthorized mutations and cannot confirm its own proposals", () =>
+it.effect("MCP rejects unauthorized mutations and saves authorized facts directly", () =>
   Effect.gen(function* () {
     const knowledge = yield* ProjectKnowledge.ProjectKnowledge;
     const projections = yield* ProjectionStore.ProjectionStoreV2;
@@ -587,7 +569,7 @@ it.effect("MCP rejects unauthorized mutations and cannot confirm its own proposa
     );
     const source = yield* knowledge.readSource({ projectId, threadId, messageId });
     if (source.state !== "available") return assert.fail("Expected canonical source");
-    const proposal = { ...createFact(projectId), sources: [source.source] };
+    const cited = { ...createFact(projectId), sources: [source.source] };
     const invoke = Effect.fn("test.knowledgeMcp")(function* (
       input: ProjectKnowledgeMutation,
       access: "read-only" | "full-access",
@@ -617,32 +599,16 @@ it.effect("MCP rejects unauthorized mutations and cannot confirm its own proposa
         .pipe(Stream.unwrap, Stream.runCollect, Effect.provide(dependencies));
       return results.at(-1)?.result;
     });
-    assert.deepInclude(yield* invoke(proposal, "read-only"), { code: "capability_denied" });
-    assert.deepInclude(yield* invoke(proposal, "full-access", false), {
+    assert.deepInclude(yield* invoke(cited, "read-only"), { code: "capability_denied" });
+    assert.deepInclude(yield* invoke(cited, "full-access", false), {
       code: "capability_denied",
     });
-    assert.deepEqual((yield* knowledge.list({ projectId, status: "proposed" })).facts, []);
-    assert.deepInclude(yield* invoke(proposal, "full-access"), { status: "proposed", revision: 1 });
-    assert.deepInclude(
-      yield* invoke(
-        {
-          action: "confirm",
-          commandId: CommandId.make("mcp-confirm"),
-          projectId,
-          factId: proposal.factId,
-          expectedRevision: 1,
-        },
-        "full-access",
-      ),
-      {
-        code: "invalid_request",
-        message: "A user must confirm proposed knowledge.",
-      },
-    );
+    assert.deepEqual((yield* knowledge.list({ projectId, status: "saved" })).facts, []);
+    assert.deepInclude(yield* invoke(cited, "full-access"), { status: "saved", revision: 1 });
     const observation = { ...createFact(projectId, "agent-observation"), sources: [] };
     const created = yield* invoke(observation, "full-access");
     assert.deepInclude(created, {
-      status: "proposed",
+      status: "saved",
       origin: { kind: "agent-client", sessionId: "test-client", label: "Test client" },
       sources: [],
     });
@@ -654,7 +620,7 @@ it.effect("MCP rejects unauthorized mutations and cannot confirm its own proposa
     assert.isTrue(
       recalled.facts.state === "available" &&
         recalled.facts.matches.some(
-          (match) => match.fact.id === observation.factId && match.fact.status === "proposed",
+          (match) => match.fact.id === observation.factId && match.fact.status === "saved",
         ),
     );
     const edited = yield* invoke(
@@ -713,11 +679,10 @@ it.effect("MCP rejects unauthorized mutations and cannot confirm its own proposa
         "full-access",
       ),
       {
-        status: "proposed",
+        status: "saved",
         origin: { kind: "agent-client", sessionId: "test-client", label: "Test client" },
       },
     );
-    assert.deepEqual((yield* knowledge.list({ projectId })).facts, []);
   }).pipe(Effect.provide(layerTest)),
 );
 
@@ -1291,9 +1256,9 @@ it.effect("pages delivery scrubbing through multiple messages in the same attemp
   }).pipe(Effect.provide(layerTest)),
 );
 
-it.effect.each(["confirm", "feedback"] as const)(
-  "scrubs a delivered %s revision using the citations that supplied its content",
-  (action) =>
+it.effect(
+  "scrubs a delivered feedback revision using the citations that supplied its content",
+  () =>
     Effect.gen(function* () {
       const knowledge = yield* ProjectKnowledge.ProjectKnowledge;
       const projections = yield* ProjectionStore.ProjectionStoreV2;
@@ -1317,11 +1282,12 @@ it.effect.each(["confirm", "feedback"] as const)(
       );
       const fact = yield* knowledge.mutate(
         {
-          commandId: CommandId.make(`delivered-${action}`),
+          action: "feedback",
+          helpful: true,
+          commandId: CommandId.make("delivered-feedback"),
           projectId,
           factId: created.id,
           expectedRevision: created.revision,
-          ...(action === "feedback" ? { action, helpful: true } : { action }),
         },
         { kind: "user" },
       );
@@ -2269,10 +2235,10 @@ it.effect("records every lifecycle event in order and erases bodies when forgott
     const projectId = ProjectId.make("history");
     yield* createProject(projectId);
     const agent = { kind: "agent-client", sessionId: "agent", label: "Agent" } as const;
-    const proposed = yield* knowledge.mutate({ ...createFact(projectId), sources: [] }, agent);
+    const created = yield* knowledge.mutate({ ...createFact(projectId), sources: [] }, agent);
     const edited = yield* knowledge.mutate(
       {
-        ...existing(proposed, "history-edit"),
+        ...existing(created, "history-edit"),
         action: "edit",
         content: "Cache keys include tenant and region IDs.",
         entities: ["cache"],
@@ -2282,27 +2248,22 @@ it.effect("records every lifecycle event in order and erases bodies when forgott
       },
       agent,
     );
-    const confirmed = yield* knowledge.mutate(
-      { ...existing(edited, "history-confirm"), action: "confirm" },
-      { kind: "user" },
-    );
     const helpful = yield* knowledge.mutate(
-      { ...existing(confirmed, "history-up"), action: "feedback", helpful: true },
+      { ...existing(edited, "history-up"), action: "feedback", helpful: true },
       { kind: "user" },
     );
     const unhelpful = yield* knowledge.mutate(
       { ...existing(helpful, "history-down"), action: "feedback", helpful: false },
       { kind: "user" },
     );
-    const detail = yield* knowledge.readFact({ projectId, factId: proposed.id });
+    const detail = yield* knowledge.readFact({ projectId, factId: created.id });
     assert.deepEqual(
       detail.history.map((entry) => [entry.event, entry.revision, entry.helpful]),
       [
         ["created", 1, null],
         ["edited", 2, null],
-        ["confirmed", 3, null],
-        ["feedback", 4, true],
-        ["feedback", 5, false],
+        ["feedback", 3, true],
+        ["feedback", 4, false],
       ],
     );
     assert.deepEqual(detail.history[0]?.actor, agent);
@@ -2310,7 +2271,7 @@ it.effect("records every lifecycle event in order and erases bodies when forgott
     assert.match(detail.history[1]?.contentDigest ?? "", /^[a-f0-9]{64}$/);
     assert.notEqual(detail.history[0]?.contentDigest, detail.history[1]?.contentDigest);
     assert.deepEqual(
-      detail.history.slice(3).map((entry) => [entry.trustBefore, entry.trustAfter]),
+      detail.history.slice(2).map((entry) => [entry.trustBefore, entry.trustAfter]),
       [
         [0.5, 0.55],
         [0.55, unhelpful.trust],
@@ -2322,7 +2283,7 @@ it.effect("records every lifecycle event in order and erases bodies when forgott
       { ...existing(unhelpful, "history-forget"), action: "forget" },
       { kind: "user" },
     );
-    const forgotten = yield* knowledge.readFact({ projectId, factId: proposed.id });
+    const forgotten = yield* knowledge.readFact({ projectId, factId: created.id });
     assert.equal(forgotten.history.at(-1)?.event, "forgotten");
     assert.deepEqual(
       forgotten.history.filter((entry) => entry.content !== null || entry.sources.length > 0),
@@ -2359,7 +2320,7 @@ it.effect("supersedes a fact by linking both ways and dropping the old one from
     const before = yield* knowledge.readFact({ projectId, factId: old.id });
     assert.equal(before.fact.supersededBy, replacement.id);
     assert.equal(before.fact.content, "The workspace uses npm for package installs.");
-    assert.equal(before.fact.status, "confirmed");
+    assert.equal(before.fact.status, "saved");
     assert.equal(before.supersededBy?.id, replacement.id);
     assert.equal(before.history.at(-1)?.event, "superseded");
     assert.equal(
@@ -2463,8 +2424,8 @@ it.effect("summarizes memory health by status, trust, feedback and lost sources"
       { ...createFact(projectId, "cited"), sources: [source.source] },
       { kind: "user" },
     );
-    const proposed = yield* knowledge.mutate(
-      { ...createFact(projectId, "proposed"), sources: [] },
+    const agentFact = yield* knowledge.mutate(
+      { ...createFact(projectId, "agent-fact"), sources: [] },
       {
         kind: "agent-client",
         sessionId: "agent",
@@ -2480,7 +2441,7 @@ it.effect("summarizes memory health by status, trust, feedback and lost sources"
         { kind: "user" },
       );
     yield* knowledge.mutate(
-      { ...existing(proposed, "praise"), action: "feedback", helpful: true },
+      { ...existing(agentFact, "praise"), action: "feedback", helpful: true },
       { kind: "user" },
     );
     const forgotten = yield* knowledge.mutate(createFact(projectId, "forgotten"), {
@@ -2509,8 +2470,7 @@ it.effect("summarizes memory health by status, trust, feedback and lost sources"
     );
 
     const status = yield* knowledge.status({ projectId });
-    assert.deepEqual(status.counts, { proposed: 1, confirmed: 3, superseded: 1, forgotten: 1 });
-    // cited 0.5, proposed 0.55, new 0.5, distrusted 0.2.
+    assert.deepEqual(status.counts, { saved: 4, superseded: 1, forgotten: 1 });
     assert.deepEqual(status.trustBands, { low: 1, guarded: 0, steady: 3, high: 0 });
     assert.equal(status.belowRecallThreshold, 1);
     assert.deepEqual(status.feedback, { helpful: 1, unhelpful: 3 });
diff --git a/apps/server/src/project/knowledge/ProjectKnowledge.ts b/apps/server/src/project/knowledge/ProjectKnowledge.ts
index d71e7f7d16ef0..0c5d94f58ca6c 100644
--- a/apps/server/src/project/knowledge/ProjectKnowledge.ts
+++ b/apps/server/src/project/knowledge/ProjectKnowledge.ts
@@ -123,15 +123,14 @@ export class ProjectKnowledge extends Context.Service<
       readonly afterId?: string | undefined;
     }) => Effect.Effect<string | null, ProjectKnowledgeError>;
     /**
-     * Creates an assisted proposal through the normal create path. An exact duplicate of an
-     * active fact creates nothing; a near duplicate or possible conflict is created as a pending
-     * replacement of that fact, which only takes effect when a user confirms it.
+     * Saves an automatically learned fact through the normal create path. An exact duplicate
+     * creates nothing; lexical similarities never block saving or replace an existing fact.
      */
-    readonly proposeAssisted: (
+    readonly saveAutomatic: (
       input: Extract<ProjectKnowledgeMutation, { readonly action: "create" }>,
       actor: Extract<KnowledgeOrigin, { readonly kind: "assisted" }>,
     ) => Effect.Effect<
-      { readonly outcome: "proposed" | "suggested" | "duplicate" | "existing" },
+      { readonly outcome: "saved" | "duplicate" | "existing" },
       ProjectKnowledgeError
     >;
     readonly hasAssistedRun: (
@@ -389,8 +388,7 @@ const make = Effect.gen(function* () {
     function* (input) {
       yield* requireProject(input.projectId);
       const limit = input.limit ?? 50;
-      // The lost-source filter matches the health count, which spans both active statuses.
-      const status = input.status ?? (input.sourceState === "unavailable" ? null : "confirmed");
+      const status = input.status ?? "saved";
       const includeSuperseded = input.includeSuperseded ?? false;
       const onlyIds =
         input.sourceState === "unavailable"
@@ -459,13 +457,6 @@ const make = Effect.gen(function* () {
                 });
               return replayed;
             }
-            if (input.action === "confirm" && actor.kind !== "user") {
-              return yield* new ProjectKnowledgeError({
-                reason: "user_confirmation_required",
-                projectId,
-                factId,
-              });
-            }
             if (
               input.action === "create" ||
               input.action === "edit" ||
@@ -560,7 +551,7 @@ const make = Effect.gen(function* () {
                 kind: input.kind,
                 sources: input.sources,
                 origin: actor,
-                status: actor.kind === "user" ? "confirmed" : "proposed",
+                status: "saved",
                 revision: 1,
                 trust: 0.5,
                 retrievalCount: 0,
@@ -601,7 +592,7 @@ const make = Effect.gen(function* () {
                 kind: input.kind,
                 sources: input.sources,
                 origin: actor,
-                status: actor.kind === "user" ? "confirmed" : "proposed",
+                status: "saved",
                 revision: 1,
                 trust: 0.5,
                 retrievalCount: 0,
@@ -622,10 +613,7 @@ const make = Effect.gen(function* () {
                 });
               if (existing.status === "forgotten")
                 return yield* new ProjectKnowledgeError({ reason: "forgotten", projectId, factId });
-              if (
-                existing.supersededBy != null &&
-                (input.action === "edit" || input.action === "confirm")
-              )
+              if (existing.supersededBy != null && input.action === "edit")
                 return yield* new ProjectKnowledgeError({
                   reason: "superseded",
                   projectId,
@@ -649,45 +637,8 @@ const make = Effect.gen(function* () {
                     kind: input.kind,
                     sources: input.sources,
                     origin: actor,
-                    status: actor.kind !== "user" ? "proposed" : existing.status,
                   };
                   break;
-                case "confirm": {
-                  if (!(yield* sourcesAvailable(existing, true)))
-                    return yield* new ProjectKnowledgeError({
-                      reason: "invalid_source",
-                      projectId,
-                      factId,
-                    });
-                  fact = { ...fact, status: "confirmed" };
-                  // An assisted suggestion names the fact it would replace without touching it;
-                  // confirming completes that replacement, or drops the link once the target is
-                  // gone or already replaced.
-                  const target =
-                    existing.supersedes == null
-                      ? null
-                      : yield* store.find(projectId, existing.supersedes);
-                  if (target !== null && target.supersededBy !== existing.id) {
-                    if (target.status === "forgotten" || target.supersededBy != null) {
-                      fact = { ...fact, supersedes: null };
-                    } else {
-                      const replaced: ProjectFact = {
-                        ...target,
-                        revision: target.revision + 1,
-                        updatedAt: now,
-                        supersededBy: existing.id,
-                      };
-                      if (!(yield* store.update(replaced, target.revision)))
-                        return yield* new ProjectKnowledgeError({
-                          reason: "conflict",
-                          projectId,
-                          factId: target.id,
-                        });
-                      yield* history(replaced, "superseded", { relatedFactId: existing.id });
-                    }
-                  }
-                  break;
-                }
                 case "forget":
                   fact = {
                     ...fact,
@@ -712,9 +663,6 @@ const make = Effect.gen(function* () {
                 case "edit":
                   yield* history(fact, "edited");
                   break;
-                case "confirm":
-                  yield* history(fact, "confirmed");
-                  break;
                 case "forget":
                   yield* store.eraseHistory(projectId, factId);
                   yield* store.scrubDeliveries(projectId, factId);
@@ -923,8 +871,6 @@ const make = Effect.gen(function* () {
           (fact): fact is typeof ActiveProjectFact.Type =>
             fact.status !== "forgotten" &&
             fact.content !== null &&
-            // Model-extracted text reaches agents only after a user confirms it.
-            !(fact.status === "proposed" && fact.origin.kind === "assisted") &&
             (entities.length === 0 ||
               (input.entityMatch === "all"
                 ? entities.every((entity) =>
@@ -1351,28 +1297,18 @@ const make = Effect.gen(function* () {
       [fact.id, fact.supersedes, fact.supersededBy].filter((id): id is ProjectFactId => id != null),
     );
   });
-  const proposeAssisted: ProjectKnowledge["Service"]["proposeAssisted"] = Effect.fn(
-    "ProjectKnowledge.proposeAssisted",
+  const saveAutomatic: ProjectKnowledge["Service"]["saveAutomatic"] = Effect.fn(
+    "ProjectKnowledge.saveAutomatic",
   )(function* (input, actor) {
     return yield* sql
       .withTransaction(
         Effect.gen(function* () {
-          // A retry reuses the same fact id, so a forgotten or reviewed proposal stays as it is.
           if ((yield* store.find(input.projectId, input.factId)) !== null)
             return { outcome: "existing" as const };
-          const hints = yield* hintsForContent(input.projectId, input.content, [input.factId]);
-          if (hints.some((hint) => hint.kind === "duplicate"))
+          if (yield* store.hasDuplicateContent(input.projectId, input.content))
             return { outcome: "duplicate" as const };
-          const created = yield* mutate(input, actor);
-          const target = hints[0];
-          if (target === undefined) return { outcome: "proposed" as const };
-          if (!(yield* store.update({ ...created, supersedes: target.factId }, created.revision)))
-            return yield* new ProjectKnowledgeError({
-              reason: "conflict",
-              projectId: input.projectId,
-              factId: input.factId,
-            });
-          return { outcome: "suggested" as const };
+          yield* mutate(input, actor);
+          return { outcome: "saved" as const };
         }),
       )
       .pipe(Effect.mapError(storageError(input.projectId)));
@@ -1446,15 +1382,19 @@ const make = Effect.gen(function* () {
       return yield* Effect.gen(function* () {
         const counts = yield* store.statusCounts(input.projectId, RECALL_THRESHOLD);
         const assistedLearning = yield* store.assistedStatus(input.projectId);
-        const cited = yield* store.citedFacts(input.projectId, null, false, SOURCE_CHECK_LIMIT + 1);
+        const cited = yield* store.citedFacts(
+          input.projectId,
+          "saved",
+          false,
+          SOURCE_CHECK_LIMIT + 1,
+        );
         const unavailable = yield* unavailableSources(
           input.projectId,
           cited.slice(0, SOURCE_CHECK_LIMIT),
         );
         return {
           counts: {
-            proposed: counts.proposed,
-            confirmed: counts.confirmed,
+            saved: counts.saved,
             superseded: counts.superseded,
             forgotten: counts.forgotten,
           },
@@ -1494,7 +1434,7 @@ const make = Effect.gen(function* () {
     markDeliveryUncertain,
     readSentDeliveries,
     scrubDeliveryPage,
-    proposeAssisted,
+    saveAutomatic,
     hasAssistedRun,
     recordAssistedRun,
   });
diff --git a/apps/server/src/project/knowledge/assistedLearningSchedule.ts b/apps/server/src/project/knowledge/assistedLearningSchedule.ts
index e364a33321891..a23cbeb3478f1 100644
--- a/apps/server/src/project/knowledge/assistedLearningSchedule.ts
+++ b/apps/server/src/project/knowledge/assistedLearningSchedule.ts
@@ -13,11 +13,6 @@ import * as Schema from "effect/Schema";
 
 import type { PendingOrchestrationEffectV2 } from "../../orchestration-v2/EffectOutbox.ts";
 
-/**
- * The extraction job for one settled run, or nothing when the project has not opted in. The
- * effect id is derived from the run, so a second finalization of the same run coalesces into
- * the job already queued.
- */
 export function assistedLearningEffects(input: {
   readonly settings: ServerSettings;
   readonly projectId: ProjectId;
diff --git a/apps/server/src/project/knowledge/knowledgeConflicts.ts b/apps/server/src/project/knowledge/knowledgeConflicts.ts
index 1864b2c581928..f2a78cc0f97d5 100644
--- a/apps/server/src/project/knowledge/knowledgeConflicts.ts
+++ b/apps/server/src/project/knowledge/knowledgeConflicts.ts
@@ -26,7 +26,8 @@ function containsStateChangeCue(text: string): boolean {
   return STATE_CHANGE_CUES.some((cue) => lower.includes(cue));
 }
 
-const normalize = (text: string) => text.split(/\s+/).filter(Boolean).join(" ").toLowerCase();
+export const normalizeFactContent = (text: string) =>
+  text.split(/\s+/).filter(Boolean).join(" ").toLowerCase();
 
 /** Dice coefficient over the same tokens recall uses. */
 function lexicalSimilarity(left: string, right: string): number {
@@ -53,7 +54,7 @@ export function conflictHints(
       revision: candidate.revision,
       excerpt: candidate.content.slice(0, 320),
     };
-    if (normalize(content) === normalize(candidate.content)) {
+    if (normalizeFactContent(content) === normalizeFactContent(candidate.content)) {
       hints.push({ ...base, kind: "duplicate", similarity: 1 });
       continue;
     }
diff --git a/apps/server/src/project/knowledge/persistence.ts b/apps/server/src/project/knowledge/persistence.ts
index 7dbb553daf582..f36f5d59345d3 100644
--- a/apps/server/src/project/knowledge/persistence.ts
+++ b/apps/server/src/project/knowledge/persistence.ts
@@ -11,6 +11,7 @@ import {
   ProjectFact,
   ProjectFactHistoryEntry,
   ProjectFactId,
+  ProjectFactStatus,
   ProjectId,
   ProjectKnowledgeDelivery,
   ProjectKnowledgeDeliveryBudget,
@@ -23,6 +24,7 @@ import {
 import * as Effect from "effect/Effect";
 import * as Schema from "effect/Schema";
 import * as SqlClient from "effect/sql/SqlClient";
+import { normalizeFactContent } from "./knowledgeConflicts.ts";
 
 const encodeDeliveryBudget = Schema.encodeEffect(
   Schema.fromJsonString(ProjectKnowledgeDeliveryBudget),
@@ -56,7 +58,7 @@ const encodeSearchReceipt = Schema.encodeSync(Schema.fromJsonString(SearchReceip
 const FactRow = Schema.Struct({
   ...ActiveProjectFact.fields,
   content: Schema.NullOr(Schema.String),
-  status: Schema.Literals(["proposed", "confirmed", "forgotten"]),
+  status: ProjectFactStatus,
   forgottenAt: Schema.NullOr(Schema.String),
   entities: Schema.fromJsonString(ActiveProjectFact.fields.entities),
   tags: Schema.fromJsonString(ActiveProjectFact.fields.tags),
@@ -172,8 +174,7 @@ export const makeStore = Effect.gen(function* () {
   });
   const list = Effect.fn("ProjectKnowledgeStore.list")(function* (
     projectId: ProjectId,
-    // null lists proposed and confirmed together.
-    status: ProjectFact["status"] | null,
+    status: ProjectFact["status"],
     afterId: string | null,
     limit: number,
     includeSuperseded: boolean,
@@ -184,7 +185,7 @@ export const makeStore = Effect.gen(function* () {
     return yield* sql`
       SELECT ${columns} FROM project_knowledge_facts
       WHERE project_id = ${projectId}
-        AND ${status === null ? sql`status IN ('confirmed', 'proposed')` : sql`status = ${status}`}
+        AND status = ${status}
         AND (${afterId} IS NULL OR id > ${afterId})
         AND (${includeSuperseded ? 1 : 0} = 1 OR status = 'forgotten' OR superseded_by_id IS NULL)
         AND ${onlyIds === null ? sql`1 = 1` : sql.in("id", onlyIds)}
@@ -272,7 +273,7 @@ export const makeStore = Effect.gen(function* () {
     minTrust: number,
     includeSuperseded: boolean,
   ) {
-    const recallable = sql`f.status IN ('confirmed', 'proposed') AND f.trust >= ${minTrust}
+    const recallable = sql`f.status = 'saved' AND f.trust >= ${minTrust}
       AND (${includeSuperseded ? 1 : 0} = 1 OR f.superseded_by_id IS NULL)`;
     const ftsQuery = terms
       .map((term) => `"${term.replaceAll('"', '""')}"${Buffer.byteLength(term) >= 4 ? "*" : ""}`)
@@ -304,6 +305,26 @@ export const makeStore = Effect.gen(function* () {
       partial: ids.length > 1000,
     };
   });
+  const hasDuplicateContent = Effect.fn("ProjectKnowledgeStore.hasDuplicateContent")(function* (
+    projectId: ProjectId,
+    content: string,
+  ) {
+    type ContentRow = { readonly id: ProjectFactId; readonly content: string };
+    const normalized = normalizeFactContent(content);
+    let afterId: ProjectFactId | null = null;
+    // Automatic saves need exact matching across all current facts, including text that the
+    // lexical tokenizer cannot index. Page only content so this does not hydrate every fact.
+    while (true) {
+      const rows: ReadonlyArray<ContentRow> = yield* sql<ContentRow>`
+        SELECT id, content FROM project_knowledge_facts
+        WHERE project_id = ${projectId} AND status = 'saved' AND superseded_by_id IS NULL
+          AND content IS NOT NULL AND id > ${afterId ?? ""}
+        ORDER BY id ASC LIMIT 100`;
+      if (rows.some((row) => normalizeFactContent(row.content) === normalized)) return true;
+      if (rows.length < 100) return false;
+      afterId = rows.at(-1)!.id;
+    }
+  });
   // Lexical neighbours for duplicate and conflict hints; bounded so a write stays cheap.
   const similar = Effect.fn("ProjectKnowledgeStore.similar")(function* (
     projectId: ProjectId,
@@ -315,7 +336,7 @@ export const makeStore = Effect.gen(function* () {
     const rows = yield* sql<{ id: ProjectFactId }>`
       SELECT f.id FROM project_knowledge_fts JOIN project_knowledge_facts f ON f.rowid = project_knowledge_fts.rowid
       WHERE project_knowledge_fts MATCH ${ftsQuery} AND f.project_id = ${projectId}
-        AND f.status IN ('confirmed', 'proposed') AND f.superseded_by_id IS NULL
+        AND f.status = 'saved' AND f.superseded_by_id IS NULL
         AND ${excludeIds.length === 0 ? sql`1 = 1` : sql`f.id NOT IN ${sql.in(excludeIds)}`}
       ORDER BY bm25(project_knowledge_fts) ASC, f.id ASC LIMIT 40`;
     return yield* findMany(
@@ -365,7 +386,6 @@ export const makeStore = Effect.gen(function* () {
     factId: ProjectFactId,
     revision: number,
   ) {
-    // Confirmation and feedback change a revision without replacing its content or citations.
     const rows = yield* sql<{ readonly sources: string }>`
       SELECT sources_json AS sources FROM project_knowledge_fact_history
       WHERE project_id = ${projectId} AND fact_id = ${factId} AND revision <= ${revision}
@@ -378,8 +398,7 @@ export const makeStore = Effect.gen(function* () {
     recallThreshold: number,
   ) {
     const [facts] = yield* sql<{
-      readonly proposed: number;
-      readonly confirmed: number;
+      readonly saved: number;
       readonly superseded: number;
       readonly forgotten: number;
       readonly low: number;
@@ -391,14 +410,11 @@ export const makeStore = Effect.gen(function* () {
     }>`
       WITH recallable AS (
         SELECT trust FROM project_knowledge_facts
-        WHERE project_id = ${projectId} AND status IN ('confirmed', 'proposed')
+        WHERE project_id = ${projectId} AND status = 'saved'
           AND superseded_by_id IS NULL
       )
       SELECT
-        (SELECT COUNT(*) FROM project_knowledge_facts WHERE project_id = ${projectId}
-          AND status = 'proposed' AND superseded_by_id IS NULL) AS proposed,
-        (SELECT COUNT(*) FROM project_knowledge_facts WHERE project_id = ${projectId}
-          AND status = 'confirmed' AND superseded_by_id IS NULL) AS confirmed,
+        (SELECT COUNT(*) FROM recallable) AS saved,
         (SELECT COUNT(*) FROM project_knowledge_facts WHERE project_id = ${projectId}
           AND status != 'forgotten' AND superseded_by_id IS NOT NULL) AS superseded,
         (SELECT COUNT(*) FROM project_knowledge_facts WHERE project_id = ${projectId}
@@ -422,14 +438,14 @@ export const makeStore = Effect.gen(function* () {
   /** Citations of active facts, for whole-project source checks. */
   const citedFacts = Effect.fn("ProjectKnowledgeStore.citedFacts")(function* (
     projectId: ProjectId,
-    status: ProjectFact["status"] | null,
+    status: ProjectFact["status"],
     includeSuperseded: boolean,
     limit: number,
   ) {
     const rows = yield* sql<{ readonly id: ProjectFactId; readonly sources: string }>`
       SELECT id, sources_json AS sources FROM project_knowledge_facts
       WHERE project_id = ${projectId}
-        AND ${status === null ? sql`status IN ('confirmed', 'proposed')` : sql`status = ${status}`}
+        AND status = ${status}
         AND (${includeSuperseded ? 1 : 0} = 1 OR superseded_by_id IS NULL)
         AND sources_json LIKE '%"message"%'
       ORDER BY id ASC LIMIT ${limit}`;
@@ -443,7 +459,7 @@ export const makeStore = Effect.gen(function* () {
   ) {
     if (ids.length === 0) return;
     yield* sql`UPDATE project_knowledge_facts SET retrieval_count = retrieval_count + 1
-      WHERE project_id = ${projectId} AND status IN ('confirmed', 'proposed') AND ${sql.in("id", ids)}`;
+      WHERE project_id = ${projectId} AND status = 'saved' AND ${sql.in("id", ids)}`;
   });
   // A sent record is final: a later re-preparation of the same attempt cannot rewrite what
   // the provider already received. Only a preparation writes excerpts; a status change keeps the
@@ -680,10 +696,10 @@ export const makeStore = Effect.gen(function* () {
     run: AssistedLearningRun,
   ) {
     yield* sql`INSERT INTO project_knowledge_assisted_runs (
-      run_id, project_id, thread_id, run_outcome, result, reason, proposed, suggested, completed_at
+      run_id, project_id, thread_id, run_outcome, result, reason, saved, completed_at
     ) VALUES (
       ${run.runId}, ${projectId}, ${run.threadId}, ${run.runOutcome}, ${run.result}, ${run.reason},
-      ${run.proposed}, ${run.suggested}, ${run.at}
+      ${run.saved}, ${run.at}
     ) ON CONFLICT (run_id) DO NOTHING`;
   });
   const assistedStatus = Effect.fn("ProjectKnowledgeStore.assistedStatus")(function* (
@@ -691,14 +707,10 @@ export const makeStore = Effect.gen(function* () {
   ) {
     const runs = yield* sql`
       SELECT thread_id AS "threadId", run_id AS "runId", run_outcome AS "runOutcome", result,
-        reason, proposed, suggested, completed_at AS at
+        reason, saved, completed_at AS at
       FROM project_knowledge_assisted_runs WHERE project_id = ${projectId}
       ORDER BY completed_at DESC, run_id DESC LIMIT 1`.pipe(Effect.flatMap(decodeAssistedRuns));
-    const [counts] = yield* sql<{ readonly suggested: number }>`
-      SELECT COUNT(*) AS suggested FROM project_knowledge_facts
-      WHERE project_id = ${projectId} AND status = 'proposed' AND superseded_by_id IS NULL
-        AND json_extract(origin_json, '$.kind') = 'assisted'`;
-    return { suggested: counts?.suggested ?? 0, lastRun: runs[0] ?? null };
+    return { lastRun: runs[0] ?? null };
   });
   return {
     assistedRunRecorded,
@@ -714,6 +726,7 @@ export const makeStore = Effect.gen(function* () {
     recordSearch,
     candidates,
     similar,
+    hasDuplicateContent,
     recordHistory,
     eraseHistory,
     history,
diff --git a/apps/server/src/provider/agentContextBudget.json b/apps/server/src/provider/agentContextBudget.json
index 94ed370fc934d..7767a8293f590 100644
--- a/apps/server/src/provider/agentContextBudget.json
+++ b/apps/server/src/provider/agentContextBudget.json
@@ -1,5 +1,5 @@
 {
-  "claudeSystemAppendBytes": 8532,
-  "codexAdditionalContextBytes": 10258,
-  "mcpToolListJsonBytes": 66447
+  "claudeSystemAppendBytes": 8591,
+  "codexAdditionalContextBytes": 10317,
+  "mcpToolListJsonBytes": 66386
 }
diff --git a/apps/web/src/components/files/FileCodeRelationships.browser.test.tsx b/apps/web/src/components/files/FileCodeRelationships.browser.test.tsx
index 9c7f3be0265f1..e37fa222b9b8d 100644
--- a/apps/web/src/components/files/FileCodeRelationships.browser.test.tsx
+++ b/apps/web/src/components/files/FileCodeRelationships.browser.test.tsx
@@ -305,7 +305,7 @@ const fact = (id: string, entities: ReadonlyArray<string>, content: string): Pro
   entities: [...entities],
   tags: [],
   sources: [],
-  status: "confirmed",
+  status: "saved",
   forgottenAt: null,
 });
 
diff --git a/apps/web/src/components/files/FileCodeRelationships.tsx b/apps/web/src/components/files/FileCodeRelationships.tsx
index 4d12c72b55ef3..b97f375d32fc4 100644
--- a/apps/web/src/components/files/FileCodeRelationships.tsx
+++ b/apps/web/src/components/files/FileCodeRelationships.tsx
@@ -511,7 +511,6 @@ function DecisionPages(props: {
             <li key={fact.id} className="flex flex-col gap-0.5 rounded-sm px-2 py-1 text-xs">
               <span className="flex items-baseline gap-2 text-2xs text-muted-foreground">
                 <span className="font-medium text-foreground">{fact.kind}</span>
-                {fact.status === "proposed" ? <span>proposed</span> : null}
                 <span className="min-w-0 flex-1 truncate">
                   {match.by === "path"
                     ? `matched by path ${match.label}`
diff --git a/apps/web/src/components/projectKnowledge/AssistedLearningSection.tsx b/apps/web/src/components/projectKnowledge/AssistedLearningSection.tsx
index 349bbc2f7993b..c224e730de43a 100644
--- a/apps/web/src/components/projectKnowledge/AssistedLearningSection.tsx
+++ b/apps/web/src/components/projectKnowledge/AssistedLearningSection.tsx
@@ -14,7 +14,6 @@ import { useEnvironmentScope } from "../../state/session";
 import { useAtomCommand } from "../../state/use-atom-command";
 import { Switch } from "../ui/switch";
 
-/** The project's opt-in for model-suggested facts and what the last run did with it. */
 export function AssistedLearningSection(props: {
   readonly environmentId: EnvironmentId;
   readonly projectId: ProjectId;
@@ -49,19 +48,19 @@ export function AssistedLearningSection(props: {
     if (result._tag === "Failure") setError(String(squashAtomCommandFailure(result)));
   };
   return (
-    <section aria-label="Assisted learning" className="flex flex-col gap-1 rounded-md border p-3">
+    <section aria-label="Automatic memory" className="flex flex-col gap-1 rounded-md border p-3">
       <div className="flex items-center justify-between gap-3 font-medium text-sm">
-        <span>Suggest facts from finished runs</span>
+        <span>Save facts from finished runs</span>
         <Switch
-          aria-label="Suggest facts from finished runs"
+          aria-label="Save facts from finished runs"
           checked={enabled}
           disabled={!canWrite || busy}
           onCheckedChange={(checked) => void toggle(checked)}
         />
       </div>
       <p className="text-xs text-muted-foreground">
-        After each run in this project, your text generation model reads that run and proposes up to
-        three cited facts for you to review. Each run costs one model call.
+        After each run in this project, your text generation model reads that run and saves up to
+        three cited facts. Each run costs one model call.
       </p>
       {lastRun ? <p className="text-xs">{assistedLearningRunLabel(lastRun)}</p> : null}
       {!canWrite ? (
diff --git a/apps/web/src/components/projectKnowledge/ProjectFactDetail.tsx b/apps/web/src/components/projectKnowledge/ProjectFactDetail.tsx
index bda4960a163a1..e6e1e3a630c48 100644
--- a/apps/web/src/components/projectKnowledge/ProjectFactDetail.tsx
+++ b/apps/web/src/components/projectKnowledge/ProjectFactDetail.tsx
@@ -227,8 +227,7 @@ export function ProjectFactHistory(props: {
 export function KnowledgeHealthStrip(props: { readonly status: ProjectKnowledgeStatus }) {
   const { status } = props;
   const counters = [
-    ["Confirmed", status.counts.confirmed],
-    ["Proposed", status.counts.proposed],
+    ["Saved", status.counts.saved],
     ["Superseded", status.counts.superseded],
     ["Below recall threshold", status.belowRecallThreshold],
     ["Retrievals", status.retrievals],
diff --git a/apps/web/src/components/projectKnowledge/ProjectMemoryFactList.tsx b/apps/web/src/components/projectKnowledge/ProjectMemoryFactList.tsx
index b3fc28e1c3c01..52929e059c935 100644
--- a/apps/web/src/components/projectKnowledge/ProjectMemoryFactList.tsx
+++ b/apps/web/src/components/projectKnowledge/ProjectMemoryFactList.tsx
@@ -80,9 +80,9 @@ export function ProjectMemoryFactList(props: {
                     <Badge size="sm" variant={status.variant}>
                       {status.label}
                     </Badge>
-                    {fact.origin.kind === "assisted" && fact.status === "proposed" ? (
+                    {fact.origin.kind === "assisted" ? (
                       <Badge size="sm" variant="outline">
-                        Suggested
+                        Automatic
                       </Badge>
                     ) : null}
                     <span className="text-muted-foreground">{fact.kind}</span>
diff --git a/apps/web/src/components/projectKnowledge/ProjectMemoryInspector.tsx b/apps/web/src/components/projectKnowledge/ProjectMemoryInspector.tsx
index c3bfcb58111fb..0d79b02046e39 100644
--- a/apps/web/src/components/projectKnowledge/ProjectMemoryInspector.tsx
+++ b/apps/web/src/components/projectKnowledge/ProjectMemoryInspector.tsx
@@ -22,7 +22,7 @@ import { useEnvironmentQuery } from "../../state/query";
 import { projectKnowledge } from "../../state/projectKnowledge";
 import { ProjectFactHistory, ProjectFactRelations } from "./ProjectFactDetail";
 import type { EditableFact } from "./ProjectFactEditor";
-import { factStatusBadge, isPendingReplacement, shortFactId } from "./projectMemoryLabels";
+import { factStatusBadge, shortFactId } from "./projectMemoryLabels";
 
 export type SourceRef = {
   readonly threadId: KnowledgeMessageRef["threadId"];
@@ -95,11 +95,7 @@ export function FactInspector(props: {
   readonly canMutate: boolean;
   readonly busy: boolean;
   readonly onEdit: (fact: EditableFact, supersede: boolean) => void;
-  readonly onRevise: (
-    fact: ProjectFact,
-    action: "confirm" | "forget" | "feedback",
-    helpful?: boolean,
-  ) => void;
+  readonly onRevise: (fact: ProjectFact, action: "forget" | "feedback", helpful?: boolean) => void;
   readonly onOpenSource: (source: KnowledgeMessageRef) => void;
   readonly onOpenFact: (factId: ProjectFactId) => void;
   readonly onBackToResults: () => void;
@@ -156,9 +152,7 @@ export function FactInspector(props: {
             <Badge variant="secondary">Superseded by {shortFactId(fact.supersededBy)}</Badge>
           ) : null}
           {fact.supersedes ? (
-            <Badge variant="outline">
-              {`${isPendingReplacement(fact) ? "Would replace" : "Supersedes"} ${shortFactId(fact.supersedes)}`}
-            </Badge>
+            <Badge variant="outline">Supersedes {shortFactId(fact.supersedes)}</Badge>
           ) : null}
         </div>
       </div>
@@ -174,19 +168,9 @@ export function FactInspector(props: {
           ))}
         </ul>
       ) : null}
-      {isPendingReplacement(fact) && fact.supersedes ? (
-        <p className="text-sm">
-          Confirming replaces fact {shortFactId(fact.supersedes)}; forgetting leaves it as it is.
-        </p>
-      ) : null}
       {fact.status !== "forgotten" ? (
         <div className="flex flex-col gap-2">
           <div className="flex flex-wrap gap-2">
-            {fact.status === "proposed" && !superseded ? (
-              <Button disabled={locked} size="sm" onClick={() => props.onRevise(fact, "confirm")}>
-                Confirm
-              </Button>
-            ) : null}
             {/* The server rejects revisions of a superseded fact; its replacement is the one to change. */}
             {!superseded ? (
               <Button
diff --git a/apps/web/src/components/projectKnowledge/ProjectMemoryPanel.browser.test.tsx b/apps/web/src/components/projectKnowledge/ProjectMemoryPanel.browser.test.tsx
index 764385566c0f1..27beaaf859d62 100644
--- a/apps/web/src/components/projectKnowledge/ProjectMemoryPanel.browser.test.tsx
+++ b/apps/web/src/components/projectKnowledge/ProjectMemoryPanel.browser.test.tsx
@@ -128,7 +128,7 @@ const fact: ProjectFact = {
   tags: [],
   kind: "constraint",
   sources: [source],
-  status: "proposed",
+  status: "saved",
   revision: 1,
   trust: 0.5,
   retrievalCount: 0,
@@ -137,7 +137,7 @@ const fact: ProjectFact = {
   forgottenAt: null,
 };
 const healthStatus: ProjectKnowledgeStatus = {
-  counts: { proposed: 1, confirmed: 4, superseded: 1, forgotten: 0 },
+  counts: { saved: 5, superseded: 1, forgotten: 0 },
   trustBands: { low: 0, guarded: 1, steady: 3, high: 1 },
   recallThreshold: 0.3,
   belowRecallThreshold: 0,
@@ -268,7 +268,6 @@ it.each([false, true])("lists only real citations under user origin (cited: %s)"
   const asserted: ProjectFact = {
     ...fact,
     origin: { kind: "user" },
-    status: "confirmed",
     sources: [{ kind: "user" }, ...(cited ? [source] : [])],
   };
   state.list = Atom.make(AsyncResult.success({ facts: [asserted], nextCursor: null }));
@@ -348,38 +347,8 @@ it("refreshes the inspected fact and uses its latest revision for the next mutat
   currentFact = { ...fact, content: "Updated on another client.", revision: 2 };
   await page.getByRole("button", { name: "Refresh", exact: true }).click();
   await expect.element(inspector().getByText(currentFact.content!)).toBeVisible();
-  await inspector().getByRole("button", { name: "Confirm", exact: true }).click();
-  expect(state.mutate).toHaveBeenCalledWith({
-    environmentId: scope.environmentId,
-    input: expect.objectContaining({ action: "confirm", expectedRevision: 2 }),
-  });
-});
-
-it("refreshes confirmed fact detail even after it leaves the proposed results", async () => {
-  const registry = await setup();
-  let currentFact: ProjectFact = fact;
-  state.detail = Atom.make(() => AsyncResult.success(factDetail(currentFact)));
-  state.list = Atom.make(() =>
-    AsyncResult.success({
-      facts: currentFact.status === "proposed" ? [currentFact] : [],
-      nextCursor: null,
-    }),
-  );
-  state.mutate.mockImplementation(async () => {
-    currentFact = { ...fact, status: "confirmed", revision: 2 };
-    return { _tag: "Success", value: currentFact };
-  });
-  await using screen = await mount(registry);
-  await openMemory(screen);
-  await page.getByRole("button", { name: "Proposed", exact: true }).click();
-  await factRow(fact.content!).click();
-  await inspector().getByRole("button", { name: "Confirm", exact: true }).click();
-  await expect.element(inspector().getByRole("heading", { name: /Confirmed/ })).toBeVisible();
-  await expect
-    .element(inspector().getByRole("button", { name: "Confirm", exact: true }))
-    .not.toBeInTheDocument();
   await inspector().getByRole("button", { name: "Helpful", exact: true }).click();
-  expect(state.mutate).toHaveBeenLastCalledWith({
+  expect(state.mutate).toHaveBeenCalledWith({
     environmentId: scope.environmentId,
     input: expect.objectContaining({ action: "feedback", expectedRevision: 2 }),
   });
@@ -710,7 +679,7 @@ it("stacks list and inspector on narrow screens and returns focus to the inspect
   await using screen = await mount(registry);
   await openMemory(screen);
   await factRow(fact.content!).click();
-  await expect.element(inspector().getByRole("heading", { name: /Proposed/ })).toHaveFocus();
+  await expect.element(inspector().getByRole("heading", { name: /Saved/ })).toHaveFocus();
   // Hidden columns leave the accessibility tree, so only one of the two is reachable.
   await expect.element(factList()).not.toBeInTheDocument();
   await page.getByRole("button", { name: "Back to list" }).click();
@@ -788,7 +757,7 @@ it("disables writes for a read-only connection", async () => {
   await openMemory(screen);
   await factRow(fact.content!).click();
   await expect.element(page.getByRole("button", { name: "Forget", exact: true })).toBeDisabled();
-  await expect.element(page.getByRole("button", { name: "Confirm", exact: true })).toBeDisabled();
+  await expect.element(page.getByRole("button", { name: "Edit", exact: true })).toBeDisabled();
   await expect.element(page.getByRole("button", { name: "New fact" })).toBeDisabled();
   expect(state.mutate).not.toHaveBeenCalled();
 });
@@ -1100,7 +1069,7 @@ it("marks superseded facts read-only and filters to facts that lost their source
   const replacement = ProjectFactId.make("replacement-fact-id");
   state.list = Atom.make(
     AsyncResult.success({
-      facts: [{ ...fact, status: "confirmed", supersededBy: replacement }],
+      facts: [{ ...fact, supersededBy: replacement }],
       nextCursor: null,
     }),
   );
@@ -1127,7 +1096,7 @@ it("marks superseded facts read-only and filters to facts that lost their source
     projectId: scope.projectId,
     sourceState: "unavailable",
   });
-  await page.getByRole("button", { name: "Confirmed", exact: true }).click();
+  await page.getByRole("button", { name: "Saved", exact: true }).click();
   await expect.element(lost).toHaveAttribute("aria-pressed", "false");
   expect(state.listRequest.mock.lastCall![0].input).not.toHaveProperty("sourceState");
 });
@@ -1181,15 +1150,14 @@ it("supersedes a fact and surfaces the conflict hints the server returns", async
   await expect.element(notice).not.toBeInTheDocument();
 });
 
-it("reviews assisted suggestions with their citations and controls the project opt-in", async () => {
+it("shows automatic facts as saved, with citations and the project switch", async () => {
   const registry = await setup();
   const runId = RunId.make("run");
-  const suggestion: ProjectFact = {
+  const automatic: ProjectFact = {
     ...fact,
     id: ProjectFactId.make("assisted:run:0"),
     origin: { kind: "assisted", threadId: scope.threadId, runId, runOutcome: "failed" },
     content: "Cache keys include tenant and region IDs.",
-    supersedes: ProjectFactId.make("existing-fact"),
   };
   state.config = Atom.make({
     environment: {
@@ -1203,7 +1171,7 @@ it("reviews assisted suggestions with their citations and controls the project o
     settings: {
       ...DEFAULT_SERVER_SETTINGS,
       projectSettingsOverrides: {
-        [scope.projectId]: { defaultAutoPull: true, enableAssistedLearning: true },
+        [scope.projectId]: { defaultAutoPull: true },
       },
     },
   });
@@ -1211,56 +1179,44 @@ it("reviews assisted suggestions with their citations and controls the project o
     AsyncResult.success({
       ...healthStatus,
       assistedLearning: {
-        suggested: 1,
         lastRun: {
           threadId: scope.threadId,
           runId,
           runOutcome: "failed",
           result: "ran",
           reason: null,
-          proposed: 1,
-          suggested: 1,
+          saved: 1,
           at: fact.createdAt,
         },
       },
     } satisfies ProjectKnowledgeStatus),
   );
-  state.list = Atom.make(AsyncResult.success({ facts: [suggestion], nextCursor: null }));
+  state.list = Atom.make(AsyncResult.success({ facts: [automatic], nextCursor: null }));
   await using screen = await mount(registry);
   await openMemory(screen);
   await page.getByRole("button", { name: "Memory health" }).click();
-  const optIn = page.getByRole("switch", { name: "Suggest facts from finished runs" });
-  await expect.element(optIn).toBeChecked();
-  await expect
-    .element(page.getByText("Last run: proposed 1 (1 as updates to existing facts)."))
-    .toBeVisible();
+  const toggle = page.getByRole("switch", { name: "Save facts from finished runs" });
+  await expect.element(toggle).toBeChecked();
+  await expect.element(page.getByText("Last run: saved 1.")).toBeVisible();
 
-  await page.getByRole("button", { name: "Suggested (1)" }).click();
+  await page.getByRole("button", { name: "Automatic", exact: true }).click();
   expect(state.listRequest).toHaveBeenLastCalledWith({
     environmentId: scope.environmentId,
-    input: { projectId: scope.projectId, status: "proposed", origin: "assisted" },
+    input: { projectId: scope.projectId, status: "saved", origin: "assisted" },
   });
-  const row = factRow(suggestion.content!);
-  await expect.element(row.getByText("Suggested", { exact: true })).toBeVisible();
+  const row = factRow(automatic.content!);
+  await expect.element(row.getByText("Saved", { exact: true })).toBeVisible();
+  await expect.element(row.getByText("Automatic", { exact: true })).toBeVisible();
   await row.click();
-  await expect.element(inspector().getByText("Suggested from a run that failed")).toBeVisible();
-  await expect.element(inspector().getByText("Would replace existing")).toBeVisible();
+  await expect.element(inspector().getByText("Learned from a run that failed")).toBeVisible();
   await inspector().getByRole("button", { name: "Message evidence" }).click();
   await expect.element(page.getByText("Exact canonical evidence.")).toBeVisible();
   await page.getByRole("button", { name: "Back to fact" }).click();
-  await inspector().getByRole("button", { name: "Confirm", exact: true }).click();
-  expect(state.mutate).toHaveBeenCalledWith({
-    environmentId: scope.environmentId,
-    input: {
-      action: "confirm",
-      commandId: expect.any(String),
-      projectId: scope.projectId,
-      factId: suggestion.id,
-      expectedRevision: 1,
-    },
-  });
+  await expect
+    .element(inspector().getByRole("button", { name: "Confirm", exact: true }))
+    .not.toBeInTheDocument();
 
-  await optIn.click();
+  await toggle.click();
   // Other overrides for the project survive the write, which replaces the whole entry.
   expect(state.updateSettings).toHaveBeenCalledWith({
     environmentId: scope.environmentId,
diff --git a/apps/web/src/components/projectKnowledge/ProjectMemoryPanel.tsx b/apps/web/src/components/projectKnowledge/ProjectMemoryPanel.tsx
index 3419a9a9b2b58..7b8512b4414cf 100644
--- a/apps/web/src/components/projectKnowledge/ProjectMemoryPanel.tsx
+++ b/apps/web/src/components/projectKnowledge/ProjectMemoryPanel.tsx
@@ -101,7 +101,6 @@ function useProjectMemoryHistoryAvailable(environmentId: EnvironmentId): boolean
   return config?.environment.capabilities.projectKnowledgeHistory === true;
 }
 
-// Older servers lack assisted proposals and the origin filter.
 function useProjectMemoryAssistedAvailable(environmentId: EnvironmentId): boolean {
   const config = useAtomValue(serverEnvironment.configValueAtom(environmentId));
   return config?.environment.capabilities.projectKnowledgeAssistedLearning === true;
@@ -158,7 +157,7 @@ export function ProjectMemoryDialogHost(props: Scope) {
         <DialogHeader>
           <DialogTitle>Project memory</DialogTitle>
           <DialogDescription>
-            Review project facts and their evidence, and choose what belongs in this draft.
+            Browse saved project facts and their evidence, and choose what belongs in this draft.
           </DialogDescription>
         </DialogHeader>
         {active ? (
@@ -174,11 +173,10 @@ export function ProjectMemoryDialogHost(props: Scope) {
   );
 }
 
-type ListFilter = "confirmed" | "proposed" | "suggested" | "forgotten" | "lost";
+type ListFilter = "saved" | "automatic" | "forgotten" | "lost";
 const FILTER_LABEL: Record<Exclude<ListFilter, "lost">, string> = {
-  confirmed: "Confirmed",
-  proposed: "Proposed",
-  suggested: "Suggested",
+  saved: "Saved",
+  automatic: "Automatic",
   forgotten: "Forgotten",
 };
 
@@ -212,7 +210,7 @@ function MemoryContents(
         }
       : null,
   );
-  const [filter, setFilter] = useState<ListFilter>("confirmed");
+  const [filter, setFilter] = useState<ListFilter>("saved");
   const [afterId, setAfterId] = useState<ProjectFactId | undefined>();
   const [includeArchived, setIncludeArchived] = useState(false);
   const [includeSuperseded, setIncludeSuperseded] = useState(false);
@@ -256,12 +254,11 @@ function MemoryContents(
       environmentId: props.environmentId,
       input: {
         projectId: props.projectId,
-        // The lost-source filter spans proposed and confirmed, matching the health count.
         ...(historyAvailable && filter === "lost"
           ? { sourceState: "unavailable" as const }
-          : filter === "suggested"
-            ? { status: "proposed" as const, origin: "assisted" as const }
-            : { status: filter === "lost" ? ("confirmed" as const) : filter }),
+          : filter === "automatic"
+            ? { status: "saved" as const, origin: "assisted" as const }
+            : { status: filter === "lost" ? ("saved" as const) : filter }),
         ...(afterId ? { afterId } : {}),
         ...(showSuperseded ? { includeSuperseded: true } : {}),
       },
@@ -410,11 +407,7 @@ function MemoryContents(
     }
     refresh([input.factId, ...(result._tag === "Success" ? [result.value.id] : [])]);
   };
-  const revise = async (
-    fact: ProjectFact,
-    action: "confirm" | "forget" | "feedback",
-    helpful = true,
-  ) => {
+  const revise = async (fact: ProjectFact, action: "forget" | "feedback", helpful = true) => {
     if (action === "forget") {
       const message = `Forget this fact permanently?\n${FORGET_CONFIRM_DETAIL}`;
       const confirmed =
@@ -481,20 +474,16 @@ function MemoryContents(
   const pending = searchInput ? search.isPending : list.isPending;
   const listError = searchInput ? search.error : list.error;
   const lostCount = health.data?.unavailableSources ?? 0;
-  const suggestedCount = health.data?.assistedLearning?.suggested ?? 0;
   const filters: ReadonlyArray<ListFilter> = [
-    "confirmed",
-    "proposed",
-    ...(assistedAvailable ? (["suggested"] as const) : []),
+    "saved",
+    ...(assistedAvailable ? (["automatic"] as const) : []),
     "forgotten",
     ...(historyAvailable && (lostCount > 0 || filter === "lost") ? (["lost"] as const) : []),
   ];
   const filterLabel = (value: ListFilter) =>
     value === "lost"
       ? `Lost source (${health.data?.unavailableSourcesCoverage === "partial" ? "at least " : ""}${lostCount})`
-      : value === "suggested" && suggestedCount > 0
-        ? `Suggested (${suggestedCount})`
-        : FILTER_LABEL[value];
+      : FILTER_LABEL[value];
   const resultSummary = pending
     ? "Reading project knowledge…"
     : searchInput
diff --git a/apps/web/src/components/projectKnowledge/projectMemoryLabels.ts b/apps/web/src/components/projectKnowledge/projectMemoryLabels.ts
index a4b5614916123..15e9c774b837d 100644
--- a/apps/web/src/components/projectKnowledge/projectMemoryLabels.ts
+++ b/apps/web/src/components/projectKnowledge/projectMemoryLabels.ts
@@ -10,13 +10,8 @@ export const selectionKey = (selection: KnowledgeSelection) =>
   selection.kind === "fact" ? factRowKey(selection.factId) : messageRowKey(selection);
 
 const STATUS_BADGE = {
-  confirmed: { label: "Confirmed", variant: "success" },
-  proposed: { label: "Proposed", variant: "info" },
+  saved: { label: "Saved", variant: "success" },
   forgotten: { label: "Forgotten", variant: "secondary" },
 } as const;
 
 export const factStatusBadge = (fact: ProjectFact) => STATUS_BADGE[fact.status];
-
-// A pending assisted replacement names its target before the target is marked superseded.
-export const isPendingReplacement = (fact: ProjectFact) =>
-  fact.status === "proposed" && fact.origin.kind === "assisted" && fact.supersedes != null;
diff --git a/apps/web/src/components/settings/ZackSettings.tsx b/apps/web/src/components/settings/ZackSettings.tsx
index ed41da6afbfee..de99ffdbece92 100644
--- a/apps/web/src/components/settings/ZackSettings.tsx
+++ b/apps/web/src/components/settings/ZackSettings.tsx
@@ -63,21 +63,26 @@ function AssistedLearningRow() {
       description={
         unavailableReason ??
         (isProjectScope
-          ? "After each run in this project, your text generation model proposes up to three cited facts for review in Project memory. Each run costs one model call."
-          : "After each run, your text generation model proposes up to three cited facts for review in Project memory. Each run costs one model call. Projects can override it.")
+          ? "After each run in this project, your text generation model saves up to three cited facts to Project memory. Each run costs one model call."
+          : "After each run, your text generation model saves up to three cited facts to Project memory. Each run costs one model call. Projects can override it.")
       }
       resetAction={
-        unavailableReason === null && settings.enableAssistedLearning ? (
+        unavailableReason === null &&
+        settings.enableAssistedLearning !== DEFAULT_SERVER_SETTINGS.enableAssistedLearning ? (
           <SettingResetButton
-            label="assisted learning"
-            tooltip="Reset assisted learning to off"
-            onClick={() => updateSettings({ enableAssistedLearning: false })}
+            label="automatic memory"
+            tooltip="Reset automatic memory to on"
+            onClick={() =>
+              updateSettings({
+                enableAssistedLearning: DEFAULT_SERVER_SETTINGS.enableAssistedLearning,
+              })
+            }
           />
         ) : null
       }
       control={
         <Switch
-          aria-label="Suggest facts from finished runs"
+          aria-label="Save facts from finished runs"
           disabled={unavailableReason !== null}
           mixed={mixed}
           checked={mixed ? false : settings.enableAssistedLearning}
diff --git a/apps/web/src/components/settings/scopedSettings.test.ts b/apps/web/src/components/settings/scopedSettings.test.ts
index 7d8382aca29e9..dc54d26bc0f15 100644
--- a/apps/web/src/components/settings/scopedSettings.test.ts
+++ b/apps/web/src/components/settings/scopedSettings.test.ts
@@ -162,7 +162,7 @@ describe("scoped settings writes", () => {
       ).toMatchObject({
         serverWrites: [],
         unavailableReason:
-          "Assisted learning is not supported on Older. Update T3 Code on that environment.",
+          "Automatic memory is not supported on Older. Update T3 Code on that environment.",
       });
     },
   );
@@ -176,7 +176,7 @@ describe("scoped settings writes", () => {
     ).toMatchObject({
       serverWrites: [],
       unavailableReason:
-        "Assisted learning is supported on only some selected environments. Update T3 Code on Server, or select a supported environment.",
+        "Automatic memory is supported on only some selected environments. Update T3 Code on Server, or select a supported environment.",
     });
   });
 
diff --git a/apps/web/src/components/settings/scopedSettings.ts b/apps/web/src/components/settings/scopedSettings.ts
index d922734ad53b6..b5be265f32984 100644
--- a/apps/web/src/components/settings/scopedSettings.ts
+++ b/apps/web/src/components/settings/scopedSettings.ts
@@ -154,8 +154,8 @@ export function scopedAssistedLearningUnavailableReason(
   if (unsupported.length === 0) return null;
   const labels = unsupported.map((environment) => environment.label).join(", ");
   return unsupported.length < environments.length
-    ? `Assisted learning is supported on only some selected environments. Update T3 Code on ${labels}, or select a supported environment.`
-    : `Assisted learning is not supported on ${labels}. Update T3 Code on ${unsupported.length === 1 ? "that environment" : "those environments"}.`;
+    ? `Automatic memory is supported on only some selected environments. Update T3 Code on ${labels}, or select a supported environment.`
+    : `Automatic memory is not supported on ${labels}. Update T3 Code on ${unsupported.length === 1 ? "that environment" : "those environments"}.`;
 }
 
 /**
diff --git a/apps/web/src/components/settings/settingsSearch.ts b/apps/web/src/components/settings/settingsSearch.ts
index 9c0f5bd012a83..9b8adfbeac780 100644
--- a/apps/web/src/components/settings/settingsSearch.ts
+++ b/apps/web/src/components/settings/settingsSearch.ts
@@ -775,12 +775,10 @@ export const SETTINGS_SEARCH_ITEMS = [
   },
   {
     id: "assisted-learning",
-    title: "Suggest facts from finished runs",
+    title: "Save facts from finished runs",
     to: "/settings/labs",
     scope: "project-defaults",
-    searchTerms: [
-      "zack assisted learning project memory knowledge facts extract suggestions model",
-    ],
+    searchTerms: ["zack automatic assisted learning project memory knowledge facts extract model"],
   },
   {
     id: "conversation-summaries",
diff --git a/docs/user/composer.md b/docs/user/composer.md
index 4a608c5bd8202..536db08850670 100644
--- a/docs/user/composer.md
+++ b/docs/user/composer.md
@@ -143,10 +143,9 @@ To save a fact from the conversation, choose **Remember for this project** on a
 message. Project memory opens that message; choose **Save as
 fact** to start a fact prefilled from it and citing it.
 
-Agents can save observations for later recall. These remain marked as proposed
-until you confirm them. Confirmation is optional; proposed observations can still
-be recalled. You can edit or forget a fact, and its origin shows who recorded the
-current assertion. Forgetting removes the fact's content everywhere T3 kept it,
+Agents can save observations for later recall. They are saved immediately, with
+no approval step. You can edit or forget a fact, and its origin shows who recorded
+the current assertion. Forgetting removes the fact's content everywhere T3 kept it,
 including its earlier revisions and the copy recorded as sent with past messages.
 
 When knowledge changes, **Supersede** the old fact instead of forgetting it. The
@@ -155,23 +154,27 @@ up in recall unless you show superseded facts. After you save, T3 points out
 existing facts that look like duplicates or possible contradictions so you can
 compare them; it never merges or replaces them on its own.
 
-To have T3 suggest facts for you, turn on **Suggest facts from finished runs** under
-**Memory health** in Project memory, or in **Settings → Labs → Zack** for one project or
-every project.
-It is off by default because each finished run in that project costs one call to
-your text generation model. After a run finishes, fails, or is stopped, T3 reads
-that run's messages and proposes up to three decisions, constraints, or
-conventions, each citing the messages it came from. Suggestions appear under
-**Suggested**, are never confirmed for you, and agents do not recall them until
-you confirm them. Suggestions from a run that failed or was stopped say so. A
-suggestion that resembles an existing fact is offered as its replacement:
-confirming it supersedes the existing fact, and forgetting it leaves that fact as
-it is. A forgotten suggestion is not proposed again for the same run. Project
-memory shows what the last run did, including why it was skipped or failed.
-
-A fact's details show its history: when it was created, edited, confirmed,
-superseded, and rated helpful or unhelpful, with the trust change from each
-rating. **Memory health** counts facts by status. When facts lose their cited
+T3 also saves facts from finished runs on its own. After a run finishes, fails, or
+is stopped, your text generation model reads that run's messages and saves up to
+three decisions, constraints, or conventions, each citing the messages it came
+from. Each run costs one model call. To turn this off, use **Save facts from
+finished runs** under **Memory health** in Project memory, or in **Settings → Labs →
+Zack** for one project or every project. The **Automatic** filter lists these facts,
+and a fact from a run that failed or was stopped says so.
+
+An exact duplicate of an existing fact is skipped. Different facts are saved
+separately, even when they look similar or may conflict. Both can come up in recall
+until you supersede one.
+Agents find saved facts by searching; nothing is added to their prompts on its own.
+A fact's trust rating, age, and how often it is recalled all affect its rank, and a
+fact rated below the recall threshold stops coming up. Facts are not deleted
+automatically.
+A fact you forget is not saved again from the same run, but a later run can learn it
+again. Project memory shows what the last run did, including why it was skipped or
+failed.
+
+A fact's details show its history: when it was created, edited, superseded, and
+rated helpful or unhelpful, with the trust change from each rating. **Memory health** counts facts by status. When facts lose their cited
 source, the **Lost source** filter lists them; on mobile, select the lost-source
 count in the memory summary.
 
diff --git a/docs/user/outside-agents.md b/docs/user/outside-agents.md
index 98f122ee5c390..c0e12c91c6a42 100644
--- a/docs/user/outside-agents.md
+++ b/docs/user/outside-agents.md
@@ -97,9 +97,8 @@ Agents without OAuth support cannot connect.
 
 Connected agents can search a project's conversations and recall saved facts.
 With permission to make changes, they can record observations for later work.
-An observation keeps the agent's identity
-and remains marked as proposed until you confirm it. Use **Project memory** in
-T3 Code to inspect, edit, or forget these facts.
+An observation is saved immediately and keeps the agent's identity. Use **Project
+memory** in T3 Code to inspect, edit, or forget these facts.
 
 Agents can also query symbols, callers, references, dependencies, and likely
 change impact in a registered project's codebase. Selecting a thread uses that
diff --git a/evals/mcp-tool-use/cases/project-intelligence.ndjson b/evals/mcp-tool-use/cases/project-intelligence.ndjson
index c3bd3bd965179..1515eeece7f80 100644
--- a/evals/mcp-tool-use/cases/project-intelligence.ndjson
+++ b/evals/mcp-tool-use/cases/project-intelligence.ndjson
@@ -1,9 +1,10 @@
-{"id":"intelligence-recall-with-citations","area":"project-intelligence","kind":"bad","pattern":"prior project decision needs memory and canonical evidence","experience":"The integration audit found no intelligence tools in the model replay catalog and no task-specific intelligence guidance in the shared provider instructions.","privacy":"commit-safe","setup":{"role":"top","mocks":{"t3_knowledge_search":{"messages":{"matches":[],"nextCursor":null},"facts":{"state":"available","coverage":"complete","nextCursor":null,"matches":[{"fact":{"id":"fact-cache-ttl","projectId":"project-app","content":"Keep the cache TTL at five minutes because downstream invalidations can arrive late.","entities":["cache"],"tags":["ttl"],"kind":"decision","sources":[{"kind":"message","threadId":"thread-cache-decision","messageId":"message-cache-decision","digest":"aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa"}],"origin":{"kind":"agent-thread","threadId":"thread-cache-decision","providerInstanceId":"codex"},"revision":1,"trust":0.9,"retrievalCount":1,"createdAt":"2026-10-22T12:00:00.000Z","updatedAt":"2026-10-22T12:00:00.000Z","status":"confirmed","forgottenAt":null},"score":{"holographic":0.9,"lexical":0.9,"tokenOverlap":1,"temporalDecay":1,"reinforcement":1,"trust":0.9,"combined":0.9}}]}},"t3_knowledge_read_source":{"state":"available","source":{"kind":"message","threadId":"thread-cache-decision","messageId":"message-cache-decision","digest":"aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa"},"text":"Keep the TTL at five minutes. Our downstream invalidation feed can arrive up to four minutes late; shortening it would break our stated cache consistency contract.","title":"Cache retention decision","role":"user","createdAt":"2026-10-22T12:00:00.000Z","archived":false}}},"prompt":"For the current machine's registered project project-app, why did we keep the cache TTL at five minutes? Find the prior decision and verify its original message.","expect":{"calls":[{"tool":"t3_knowledge_search","args":{"projectId":"project-app","query":"~cache"}},{"tool":"t3_knowledge_read_source","args":{"projectId":"project-app","threadId":"thread-cache-decision","messageId":"message-cache-decision","digest":"aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa"}}],"order":["t3_knowledge_search","t3_knowledge_read_source"],"notCalls":["t3_knowledge_mutate","t3_project_code_query"],"maxCalls":{"t3_knowledge_search":1}},"source":{"kind":"integration-audit","commit":"47021ecc04f525c260159bd82f9f16e748ce41c3","path":"evals/mcp-tool-use/run.ts","finding":"The agent replay catalog omitted the knowledge, conversation summary, and project code toolkits."}}
+{"id":"intelligence-recall-with-citations","area":"project-intelligence","kind":"bad","pattern":"prior project decision needs memory and canonical evidence","experience":"The integration audit found no intelligence tools in the model replay catalog and no task-specific intelligence guidance in the shared provider instructions.","privacy":"commit-safe","setup":{"role":"top","mocks":{"t3_knowledge_search":{"messages":{"matches":[],"nextCursor":null},"facts":{"state":"available","coverage":"complete","nextCursor":null,"matches":[{"fact":{"id":"fact-cache-ttl","projectId":"project-app","content":"Keep the cache TTL at five minutes because downstream invalidations can arrive late.","entities":["cache"],"tags":["ttl"],"kind":"decision","sources":[{"kind":"message","threadId":"thread-cache-decision","messageId":"message-cache-decision","digest":"aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa"}],"origin":{"kind":"agent-thread","threadId":"thread-cache-decision","providerInstanceId":"codex"},"revision":1,"trust":0.9,"retrievalCount":1,"createdAt":"2026-10-22T12:00:00.000Z","updatedAt":"2026-10-22T12:00:00.000Z","status":"saved","forgottenAt":null},"score":{"holographic":0.9,"lexical":0.9,"tokenOverlap":1,"temporalDecay":1,"reinforcement":1,"trust":0.9,"combined":0.9}}]}},"t3_knowledge_read_source":{"state":"available","source":{"kind":"message","threadId":"thread-cache-decision","messageId":"message-cache-decision","digest":"aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa"},"text":"Keep the TTL at five minutes. Our downstream invalidation feed can arrive up to four minutes late; shortening it would break our stated cache consistency contract.","title":"Cache retention decision","role":"user","createdAt":"2026-10-22T12:00:00.000Z","archived":false}}},"prompt":"For the current machine's registered project project-app, why did we keep the cache TTL at five minutes? Find the prior decision and verify its original message.","expect":{"calls":[{"tool":"t3_knowledge_search","args":{"projectId":"project-app","query":"~cache"}},{"tool":"t3_knowledge_read_source","args":{"projectId":"project-app","threadId":"thread-cache-decision","messageId":"message-cache-decision","digest":"aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa"}}],"order":["t3_knowledge_search","t3_knowledge_read_source"],"notCalls":["t3_knowledge_mutate","t3_project_code_query"],"maxCalls":{"t3_knowledge_search":1}},"source":{"kind":"integration-audit","commit":"47021ecc04f525c260159bd82f9f16e748ce41c3","path":"evals/mcp-tool-use/run.ts","finding":"The agent replay catalog omitted the knowledge, conversation summary, and project code toolkits."}}
 {"id":"intelligence-code-impact-worktree","area":"project-intelligence","kind":"bad","pattern":"symbol impact must use the registered worktree","experience":"The integration audit found no intelligence tools in the model replay catalog and no task-specific intelligence guidance in the shared provider instructions.","privacy":"commit-safe","setup":{"role":"top","mocks":{"t3_project_code_query":{"operation":"impact","checkoutPath":"/home/user/app-worktree","matches":[],"nodes":[{"id":"symbol-cache","name":"readCache","qualifiedName":"readCache","kind":"function","language":"typescript","path":"src/cache.ts","startLine":10,"endLine":14,"startColumn":0,"endColumn":1,"signature":null,"isExported":true},{"id":"symbol-loader","name":"loadUser","qualifiedName":"loadUser","kind":"function","language":"typescript","path":"src/loader.ts","startLine":20,"endLine":24,"startColumn":0,"endColumn":1,"signature":null,"isExported":true}],"edges":[{"source":"symbol-loader","target":"symbol-cache","kind":"call","confidence":0.8,"resolvedBy":"import","provenance":"tree-sitter","evidenceGrade":"static-candidate","line":22,"column":2,"referenceName":"readCache","candidateTargets":[]}],"files":[{"path":"src/cache.ts"},{"path":"src/loader.ts"}],"unresolvedReferences":[],"selection":{"state":"matched","candidates":[{"id":"symbol-cache","name":"readCache","qualifiedName":"readCache","kind":"function","language":"typescript","path":"src/cache.ts","startLine":10,"endLine":14,"startColumn":0,"endColumn":1,"signature":null,"isExported":true}]},"sourceDigests":[{"path":"src/cache.ts","digest":"aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa","verified":true}],"stats":null,"coverage":{"state":"complete","catalogRevision":"test","relationSemantics":"static-candidates","columnEncoding":"utf16-code-units","languages":["typescript"],"indexedLanguages":["typescript"],"filesListed":2,"filesIndexed":2,"filesSkipped":0,"filesUnsupported":0,"filesFailed":0,"pendingReferences":0,"resultLimitReached":false,"limitations":["Static candidates require source inspection before claiming runtime reachability."]}}},"world":{"files":{"src/cache.ts":"export function readCache(key: string) { return cache.get(key); }\n","src/loader.ts":"import { readCache } from './cache';\nexport function loadUser(id: string) { return readCache(id); }\n"}}},"prompt":"In registered project project-app, thread thread-feature uses a worktree. What could be affected by changing readCache in src/cache.ts there? Start with the code graph; no edits.","expect":{"calls":[{"tool":"t3_project_code_query","args":{"projectId":"project-app","threadId":"thread-feature","query.operation":"impact","query.symbol.name":"readCache","query.symbol.filePath":"src/cache.ts"}}],"notCalls":["t3_knowledge_mutate","t3_knowledge_search"]},"source":{"kind":"integration-audit","commit":"47021ecc04f525c260159bd82f9f16e748ce41c3","path":"evals/mcp-tool-use/run.ts","finding":"The agent replay catalog omitted the knowledge, conversation summary, and project code toolkits."}}
 {"id": "intelligence-summary-exact-source", "area": "project-intelligence", "kind": "bad", "pattern": "a model summary is not the original decision", "experience": "The integration audit found no intelligence tools in the model replay catalog and no task-specific intelligence guidance in the shared provider instructions.", "privacy": "commit-safe", "setup": {"role": "top", "mocks": {"t3_conversation_summary_expand": {"summary": {"state": "available", "summary": {"id": "summary:bbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbb", "projectId": "project-app", "threadId": "thread-cache-decision", "sources": [{"kind": "message", "threadId": "thread-cache-decision", "messageId": "message-cache-decision", "digest": "aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa"}], "statements": [{"text": "The team agreed to keep the TTL at five minutes.", "citations": ["message-cache-decision"]}], "modelSelection": {"instanceId": "codex", "model": "gpt-5"}, "createdAt": "2026-10-22T12:00:00.000Z"}}, "sources": [{"ref": {"kind": "message", "threadId": "thread-cache-decision", "messageId": "message-cache-decision", "digest": "aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa"}, "result": {"state": "available", "source": {"kind": "message", "threadId": "thread-cache-decision", "messageId": "message-cache-decision", "digest": "aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa"}, "text": "Keep the TTL at five minutes. Our downstream invalidation feed can arrive up to four minutes late; shortening it would break our stated cache consistency contract.", "title": "Cache retention decision", "role": "user", "createdAt": "2026-10-22T12:00:00.000Z", "archived": false}}], "nextOffset": null, "nextTextOffset": null}}}, "prompt": "The handoff for project project-app, thread thread-cache-decision, includes conversation summary summary:bbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbb: 'The team agreed to keep the TTL at five minutes.' What did the user actually say? Verify the original message; do not treat the summary as a quotation.", "expect": {"calls": [{"tool": "t3_conversation_summary_expand", "args": {"projectId": "project-app", "threadId": "thread-cache-decision", "summaryId": "summary:bbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbb"}}], "textMust": ["four minutes|four-minute"], "notCalls": ["t3_knowledge_mutate", "t3_project_code_query"]}, "source": {"kind": "integration-audit", "commit": "47021ecc04f525c260159bd82f9f16e748ce41c3", "path": "evals/mcp-tool-use/run.ts", "finding": "The agent replay catalog omitted the knowledge, conversation summary, and project code toolkits."}}
-{"id":"intelligence-remote-recall-scope","area":"project-intelligence","kind":"guard","pattern":"knowledge IDs and citations stay on the selected environment","experience":"The integration audit found no intelligence tools in the model replay catalog and no task-specific intelligence guidance in the shared provider instructions.","privacy":"commit-safe","setup":{"role":"top","mocks":{"t3_knowledge_search":{"messages":{"matches":[],"nextCursor":null},"facts":{"state":"available","coverage":"complete","nextCursor":null,"matches":[{"fact":{"id":"fact-cache-ttl","projectId":"project-build","content":"Keep the cache TTL at five minutes because downstream invalidations can arrive late.","entities":["cache"],"tags":["ttl"],"kind":"decision","sources":[{"kind":"message","threadId":"thread-cache-decision","messageId":"message-cache-decision","digest":"aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa"}],"origin":{"kind":"agent-thread","threadId":"thread-cache-decision","providerInstanceId":"codex"},"revision":1,"trust":0.9,"retrievalCount":1,"createdAt":"2026-10-22T12:00:00.000Z","updatedAt":"2026-10-22T12:00:00.000Z","status":"confirmed","forgottenAt":null},"score":{"holographic":0.9,"lexical":0.9,"tokenOverlap":1,"temporalDecay":1,"reinforcement":1,"trust":0.9,"combined":0.9}}]}},"t3_knowledge_read_source":{"state":"available","source":{"kind":"message","threadId":"thread-cache-decision","messageId":"message-cache-decision","digest":"aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa"},"text":"Keep the TTL at five minutes. Our downstream invalidation feed can arrive up to four minutes late; shortening it would break our stated cache consistency contract.","title":"Cache retention decision","role":"user","createdAt":"2026-10-22T12:00:00.000Z","archived":false}}},"prompt":"Use the linked build machine (environmentId build-box), registered project project-build. Why did we keep the cache TTL at five minutes there? Verify the original cited message on that same machine.","expect":{"calls":[{"tool":"t3_knowledge_search","args":{"environmentId":"build-box","projectId":"project-build","query":"~cache"}},{"tool":"t3_knowledge_read_source","args":{"environmentId":"build-box","projectId":"project-build","threadId":"thread-cache-decision","messageId":"message-cache-decision"}}],"order":["t3_knowledge_search","t3_knowledge_read_source"],"maxCalls":{"t3_knowledge_search":1,"t3_knowledge_read_source":1},"notCalls":["t3_knowledge_mutate"]},"source":{"kind":"integration-audit","commit":"47021ecc04f525c260159bd82f9f16e748ce41c3","path":"evals/mcp-tool-use/run.ts","finding":"The agent replay catalog omitted the knowledge, conversation summary, and project code toolkits."}}
+{"id":"intelligence-remote-recall-scope","area":"project-intelligence","kind":"guard","pattern":"knowledge IDs and citations stay on the selected environment","experience":"The integration audit found no intelligence tools in the model replay catalog and no task-specific intelligence guidance in the shared provider instructions.","privacy":"commit-safe","setup":{"role":"top","mocks":{"t3_knowledge_search":{"messages":{"matches":[],"nextCursor":null},"facts":{"state":"available","coverage":"complete","nextCursor":null,"matches":[{"fact":{"id":"fact-cache-ttl","projectId":"project-build","content":"Keep the cache TTL at five minutes because downstream invalidations can arrive late.","entities":["cache"],"tags":["ttl"],"kind":"decision","sources":[{"kind":"message","threadId":"thread-cache-decision","messageId":"message-cache-decision","digest":"aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa"}],"origin":{"kind":"agent-thread","threadId":"thread-cache-decision","providerInstanceId":"codex"},"revision":1,"trust":0.9,"retrievalCount":1,"createdAt":"2026-10-22T12:00:00.000Z","updatedAt":"2026-10-22T12:00:00.000Z","status":"saved","forgottenAt":null},"score":{"holographic":0.9,"lexical":0.9,"tokenOverlap":1,"temporalDecay":1,"reinforcement":1,"trust":0.9,"combined":0.9}}]}},"t3_knowledge_read_source":{"state":"available","source":{"kind":"message","threadId":"thread-cache-decision","messageId":"message-cache-decision","digest":"aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa"},"text":"Keep the TTL at five minutes. Our downstream invalidation feed can arrive up to four minutes late; shortening it would break our stated cache consistency contract.","title":"Cache retention decision","role":"user","createdAt":"2026-10-22T12:00:00.000Z","archived":false}}},"prompt":"Use the linked build machine (environmentId build-box), registered project project-build. Why did we keep the cache TTL at five minutes there? Verify the original cited message on that same machine.","expect":{"calls":[{"tool":"t3_knowledge_search","args":{"environmentId":"build-box","projectId":"project-build","query":"~cache"}},{"tool":"t3_knowledge_read_source","args":{"environmentId":"build-box","projectId":"project-build","threadId":"thread-cache-decision","messageId":"message-cache-decision"}}],"order":["t3_knowledge_search","t3_knowledge_read_source"],"maxCalls":{"t3_knowledge_search":1,"t3_knowledge_read_source":1},"notCalls":["t3_knowledge_mutate"]},"source":{"kind":"integration-audit","commit":"47021ecc04f525c260159bd82f9f16e748ce41c3","path":"evals/mcp-tool-use/run.ts","finding":"The agent replay catalog omitted the knowledge, conversation summary, and project code toolkits."}}
 {"id":"intelligence-return-to-local-scope","area":"project-intelligence","kind":"control","pattern":"a previous remote query does not change the current environment","experience":"The integration audit found no intelligence tools in the model replay catalog and no task-specific intelligence guidance in the shared provider instructions.","privacy":"commit-safe","setup":{"role":"top","priorTurns":[{"role":"user","text":"Inspect readCache on the linked build box."},{"role":"assistant","text":"The build box has one static caller.","toolCalls":[{"tool":"t3_project_code_query","input":{"environmentId":"build-box","projectId":"project-build","query":{"operation":"impact","symbol":{"name":"readCache","filePath":"src/cache.ts"}}},"result":{"operation":"impact","checkoutPath":"/home/user/app-worktree","matches":[],"nodes":[{"id":"symbol-cache","name":"readCache","qualifiedName":"readCache","kind":"function","language":"typescript","path":"src/cache.ts","startLine":10,"endLine":14,"startColumn":0,"endColumn":1,"signature":null,"isExported":true},{"id":"symbol-loader","name":"loadUser","qualifiedName":"loadUser","kind":"function","language":"typescript","path":"src/loader.ts","startLine":20,"endLine":24,"startColumn":0,"endColumn":1,"signature":null,"isExported":true}],"edges":[{"source":"symbol-loader","target":"symbol-cache","kind":"call","confidence":0.8,"resolvedBy":"import","provenance":"tree-sitter","evidenceGrade":"static-candidate","line":22,"column":2,"referenceName":"readCache","candidateTargets":[]}],"files":[{"path":"src/cache.ts"},{"path":"src/loader.ts"}],"unresolvedReferences":[],"selection":{"state":"matched","candidates":[{"id":"symbol-cache","name":"readCache","qualifiedName":"readCache","kind":"function","language":"typescript","path":"src/cache.ts","startLine":10,"endLine":14,"startColumn":0,"endColumn":1,"signature":null,"isExported":true}]},"sourceDigests":[{"path":"src/cache.ts","digest":"aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa","verified":true}],"stats":null,"coverage":{"state":"complete","catalogRevision":"test","relationSemantics":"static-candidates","columnEncoding":"utf16-code-units","languages":["typescript"],"indexedLanguages":["typescript"],"filesListed":2,"filesIndexed":2,"filesSkipped":0,"filesUnsupported":0,"filesFailed":0,"pendingReferences":0,"resultLimitReached":false,"limitations":["Static candidates require source inspection before claiming runtime reachability."]}}}]}],"mocks":{"t3_project_code_query":{"operation":"impact","checkoutPath":"/home/user/app-worktree","matches":[],"nodes":[{"id":"symbol-cache","name":"readCache","qualifiedName":"readCache","kind":"function","language":"typescript","path":"src/cache.ts","startLine":10,"endLine":14,"startColumn":0,"endColumn":1,"signature":null,"isExported":true},{"id":"symbol-loader","name":"loadUser","qualifiedName":"loadUser","kind":"function","language":"typescript","path":"src/loader.ts","startLine":20,"endLine":24,"startColumn":0,"endColumn":1,"signature":null,"isExported":true}],"edges":[{"source":"symbol-loader","target":"symbol-cache","kind":"call","confidence":0.8,"resolvedBy":"import","provenance":"tree-sitter","evidenceGrade":"static-candidate","line":22,"column":2,"referenceName":"readCache","candidateTargets":[]}],"files":[{"path":"src/cache.ts"},{"path":"src/loader.ts"}],"unresolvedReferences":[],"selection":{"state":"matched","candidates":[{"id":"symbol-cache","name":"readCache","qualifiedName":"readCache","kind":"function","language":"typescript","path":"src/cache.ts","startLine":10,"endLine":14,"startColumn":0,"endColumn":1,"signature":null,"isExported":true}]},"sourceDigests":[{"path":"src/cache.ts","digest":"aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa","verified":true}],"stats":null,"coverage":{"state":"complete","catalogRevision":"test","relationSemantics":"static-candidates","columnEncoding":"utf16-code-units","languages":["typescript"],"indexedLanguages":["typescript"],"filesListed":2,"filesIndexed":2,"filesSkipped":0,"filesUnsupported":0,"filesFailed":0,"pendingReferences":0,"resultLimitReached":false,"limitations":["Static candidates require source inspection before claiming runtime reachability."]}}}},"prompt":"Now check the current machine instead: project project-app, thread thread-feature. Use its code graph to find the impact of readCache in src/cache.ts.","expect":{"calls":[{"tool":"t3_project_code_query","args":{"projectId":"project-app","threadId":"thread-feature","query.operation":"impact","query.symbol.name":"readCache"}}],"callsMustNot":[{"tool":"t3_project_code_query","args":{"environmentId":"build-box"}}],"maxCalls":{"t3_project_code_query":1},"notCalls":["t3_environment_link","t3_environment_links","t3_knowledge_search"]},"source":{"kind":"integration-audit","commit":"47021ecc04f525c260159bd82f9f16e748ce41c3","path":"evals/mcp-tool-use/run.ts","finding":"The agent replay catalog omitted the knowledge, conversation summary, and project code toolkits."}}
 {"id":"intelligence-ordinary-edit-control","area":"project-intelligence","kind":"control","pattern":"a self-contained edit needs no project intelligence calls","experience":"The integration audit found no intelligence tools in the model replay catalog and no task-specific intelligence guidance in the shared provider instructions.","privacy":"commit-safe","setup":{"role":"top","world":{"branch":"fix/save-label","files":{"src/labels.ts":"export const saveLabel = 'Save changes';\n"}}},"prompt":"In src/labels.ts change the exported saveLabel string from 'Save changes' to 'Save'. This is just a copy edit; do not change anything else.","expect":{"calls":[{"tool":"Edit|Write|Bash"}],"notCalls":["t3_knowledge_search","t3_knowledge_list","t3_knowledge_read_fact","t3_knowledge_status","t3_knowledge_read_source","t3_knowledge_prepare_context","t3_knowledge_mutate","t3_conversation_summary_list","t3_conversation_summary_describe","t3_conversation_summary_expand","t3_project_code_query","t3_project_file_history"]},"source":{"kind":"integration-audit","commit":"47021ecc04f525c260159bd82f9f16e748ce41c3","path":"evals/mcp-tool-use/run.ts","finding":"The agent replay catalog omitted the knowledge, conversation summary, and project code toolkits."}}
 {"id":"intelligence-native-subagent-worktree","area":"project-intelligence","kind":"guard","pattern":"native child receives project intelligence guidance","experience":"The integration audit found that Claude native children received GitHub and device guidance but not the parent project-intelligence guidance.","privacy":"commit-safe","setup":{"role":"native-subagent","mocks":{"t3_project_code_query":{"operation":"impact","checkoutPath":"/home/user/app-worktree","matches":[],"nodes":[{"id":"symbol-cache","name":"readCache","qualifiedName":"readCache","kind":"function","language":"typescript","path":"src/cache.ts","startLine":10,"endLine":14,"startColumn":0,"endColumn":1,"signature":null,"isExported":true},{"id":"symbol-loader","name":"loadUser","qualifiedName":"loadUser","kind":"function","language":"typescript","path":"src/loader.ts","startLine":20,"endLine":24,"startColumn":0,"endColumn":1,"signature":null,"isExported":true}],"edges":[{"source":"symbol-loader","target":"symbol-cache","kind":"call","confidence":0.8,"resolvedBy":"import","provenance":"tree-sitter","evidenceGrade":"static-candidate","line":22,"column":2,"referenceName":"readCache","candidateTargets":[]}],"files":[{"path":"src/cache.ts"},{"path":"src/loader.ts"}],"unresolvedReferences":[],"selection":{"state":"matched","candidates":[{"id":"symbol-cache","name":"readCache","qualifiedName":"readCache","kind":"function","language":"typescript","path":"src/cache.ts","startLine":10,"endLine":14,"startColumn":0,"endColumn":1,"signature":null,"isExported":true}]},"sourceDigests":[{"path":"src/cache.ts","digest":"aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa","verified":true}],"stats":null,"coverage":{"state":"complete","catalogRevision":"test","relationSemantics":"static-candidates","columnEncoding":"utf16-code-units","languages":["typescript"],"indexedLanguages":["typescript"],"filesListed":2,"filesIndexed":2,"filesSkipped":0,"filesUnsupported":0,"filesFailed":0,"pendingReferences":0,"resultLimitReached":false,"limitations":["Static candidates require source inspection before claiming runtime reachability."]}}},"world":{"files":{"src/cache.ts":"export function readCache(key: string) { return cache.get(key); }\n","src/loader.ts":"import { readCache } from './cache';\nexport function loadUser(id: string) { return readCache(id); }\n"},"deviceAgentAccess":false},"harness":"claude"},"prompt":"In registered project project-app, thread thread-feature uses a worktree. What could be affected by changing readCache in src/cache.ts there? Start with the code graph; no edits.","expect":{"calls":[{"tool":"t3_project_code_query","args":{"projectId":"project-app","threadId":"thread-feature","query.operation":"impact","query.symbol.name":"readCache","query.symbol.filePath":"src/cache.ts"}}],"notCalls":["t3_knowledge_mutate","t3_knowledge_search"]},"source":{"kind":"integration-audit","commit":"e747cef0ba5606bcc19f228963a84935da27f3f6","path":"apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.ts","finding":"Claude native subagents do not inherit the parent system prompt, and SubagentStart omitted the shared project-intelligence guidance."}}
 {"id":"intelligence-native-subagent-edit-control","area":"project-intelligence","kind":"control","pattern":"native child ordinary edits need no project intelligence","experience":"Project-intelligence guidance must not cause native children to retrieve history or a graph for a routine edit.","privacy":"commit-safe","setup":{"role":"native-subagent","world":{"branch":"fix/save-label","files":{"src/labels.ts":"export const saveLabel = 'Save changes';\n"},"deviceAgentAccess":false},"harness":"claude"},"prompt":"In src/labels.ts change the exported saveLabel string from 'Save changes' to 'Save'. This is just a copy edit; do not change anything else.","expect":{"calls":[{"tool":"Edit|Write|Bash"}],"notCalls":["t3_knowledge_search","t3_knowledge_list","t3_knowledge_read_fact","t3_knowledge_status","t3_knowledge_read_source","t3_knowledge_prepare_context","t3_knowledge_mutate","t3_conversation_summary_list","t3_conversation_summary_describe","t3_conversation_summary_expand","t3_project_code_query","t3_project_file_history"]},"source":{"kind":"integration-audit","commit":"e747cef0ba5606bcc19f228963a84935da27f3f6","path":"apps/server/src/orchestration-v2/Adapters/ClaudeAdapterV2.ts","finding":"Claude native subagents do not inherit the parent system prompt, and SubagentStart omitted the shared project-intelligence guidance."}}
 {"id": "intelligence-summary-source-continuation", "area": "project-intelligence", "kind": "guard", "pattern": "bounded source expansion resumes within the same source", "experience": "The LCM audit found unbounded source expansion. Bounded pages now carry both a source offset and an offset within that source, and agents must use both to recover evidence after the first page.", "privacy": "commit-safe", "source": {"kind": "integration-audit", "commit": "ef873edeb5927d63e4e40920923b629616b95bbf", "path": "apps/server/src/project/conversationSummaries/ConversationSummaries.ts", "finding": "Expansion returned complete source messages with only source-count pagination."}, "setup": {"role": "top", "priorTurns": [{"role": "user", "text": "Read the original health-check output for project-app in thread-health-check."}, {"role": "assistant", "text": "The first page stops before the result; more source text remains.", "toolCalls": [{"tool": "t3_conversation_summary_expand", "input": {"projectId": "project-app", "threadId": "thread-health-check", "summaryId": "summary:dddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddd", "offset": 0, "maxChars": 42}, "result": {"summary": {"state": "available", "summary": {"id": "summary:dddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddd", "projectId": "project-app", "threadId": "thread-health-check", "sources": [{"kind": "turn_item", "threadId": "thread-health-check", "itemId": "item-health-output", "digest": "cccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccc"}], "statements": [{"text": "The health check failed.", "citations": ["turn_item:item-health-output"]}], "modelSelection": {"instanceId": "codex", "model": "gpt-5"}, "createdAt": "2026-11-03T00:00:00.000Z"}}, "sources": [{"ref": {"kind": "turn_item", "threadId": "thread-health-check", "itemId": "item-health-output", "digest": "cccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccc"}, "result": {"state": "available", "source": {"kind": "turn_item", "threadId": "thread-health-check", "itemId": "item-health-output", "digest": "cccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccc"}, "text": "command: curl /health\nstatus: checking...\n", "title": "Health check", "role": "tool", "createdAt": "2026-11-03T00:00:00.000Z", "archived": false, "textOffset": 0, "totalChars": 91}}], "nextOffset": 0, "nextTextOffset": 42}}]}], "mocks": {"t3_conversation_summary_expand": {"summary": {"state": "available", "summary": {"id": "summary:dddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddd", "projectId": "project-app", "threadId": "thread-health-check", "sources": [{"kind": "turn_item", "threadId": "thread-health-check", "itemId": "item-health-output", "digest": "cccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccc"}], "statements": [{"text": "The health check failed.", "citations": ["turn_item:item-health-output"]}], "modelSelection": {"instanceId": "codex", "model": "gpt-5"}, "createdAt": "2026-11-03T00:00:00.000Z"}}, "sources": [{"ref": {"kind": "turn_item", "threadId": "thread-health-check", "itemId": "item-health-output", "digest": "cccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccc"}, "result": {"state": "available", "source": {"kind": "turn_item", "threadId": "thread-health-check", "itemId": "item-health-output", "digest": "cccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccccc"}, "text": "AssertionError: expected HTTP 200, received 503.\n", "title": "Health check", "role": "tool", "createdAt": "2026-11-03T00:00:00.000Z", "archived": false, "textOffset": 42, "totalChars": 91}}], "nextOffset": null, "nextTextOffset": null}}}, "prompt": "Continue reading from the returned cursor. What did the original health-check output actually report? Identify the source item.", "expect": {"calls": [{"tool": "t3_conversation_summary_expand", "args": {"projectId": "project-app", "threadId": "thread-health-check", "summaryId": "summary:dddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddd", "offset": 0, "textOffset": 42}}], "maxCalls": {"t3_conversation_summary_expand": 1}, "notCalls": ["t3_knowledge_mutate", "t3_project_code_query"], "textMust": ["503", "item-health-output"]}}
+{"id":"intelligence-remember-saves-directly","area":"project-intelligence","kind":"guard","pattern":"remembering a fact saves it with no approval step","experience":"Agents described saved facts as proposals a user had to confirm. Facts now save directly, and no confirm action exists.","privacy":"commit-safe","source":{"kind":"integration-audit","path":"packages/provider-core/src/server/orchestrationInstructions.ts","finding":"Project memory moved to a saved or forgotten lifecycle with no approval step."},"setup":{"role":"top","mocks":{"t3_knowledge_mutate":{"id":"fact-cache-ttl-saved","projectId":"project-app","content":"Keep the cache TTL at five minutes.","entities":["cache"],"tags":["ttl"],"kind":"decision","sources":[],"origin":{"kind":"agent-thread","threadId":"thread-current","providerInstanceId":"codex"},"revision":1,"trust":0.5,"retrievalCount":0,"createdAt":"2026-10-22T12:00:00.000Z","updatedAt":"2026-10-22T12:00:00.000Z","status":"saved","forgottenAt":null,"hints":[]}}},"prompt":"For the current machine's registered project project-app, remember that we keep the cache TTL at five minutes. Tell me whether other agents can use it yet or whether I still need to do anything.","expect":{"calls":[{"tool":"t3_knowledge_mutate","args":{"mutation":{"action":"create","projectId":"project-app","content":"~five minutes"}}}],"notCalls":["t3_knowledge_prepare_context"],"maxCalls":{"t3_knowledge_mutate":1},"finalTextMust":["(?i:saved|remembered|recorded|stored)"],"textMustNot":["(?i:(?:need|have|must)[^.]{0,40}\\b(?:confirm|approve|review)\\b|until (?:you|a user) (?:confirm|approve|review)|pending (?:confirmation|approval|review)|awaiting (?:confirmation|approval|review))"]}}
diff --git a/packages/client-runtime/src/projectKnowledge/index.ts b/packages/client-runtime/src/projectKnowledge/index.ts
index c0ea78e063a3e..598ff3acb56ec 100644
--- a/packages/client-runtime/src/projectKnowledge/index.ts
+++ b/packages/client-runtime/src/projectKnowledge/index.ts
@@ -154,20 +154,19 @@ export function knowledgeOriginLabel(origin: KnowledgeOrigin): string {
       return `Agent observation from ${origin.label}`;
     case "assisted":
       return origin.runOutcome === "completed"
-        ? "Suggested from a finished run"
-        : `Suggested from a run that ${origin.runOutcome === "failed" ? "failed" : "was interrupted"}`;
+        ? "Learned from a finished run"
+        : `Learned from a run that ${origin.runOutcome === "failed" ? "failed" : "was interrupted"}`;
     case "unknown":
       return "Origin not recognized by this version";
   }
 }
 
-/** One plain line for the project's latest assisted learning outcome. */
 export function assistedLearningRunLabel(run: AssistedLearningRun): string {
   switch (run.result) {
     case "ran":
-      return run.proposed === 0
-        ? `Last run: nothing proposed.${run.reason ? ` ${run.reason}` : ""}`
-        : `Last run: proposed ${run.proposed}${run.suggested > 0 ? ` (${run.suggested} as updates to existing facts)` : ""}.${run.reason ? ` ${run.reason}` : ""}`;
+      return run.saved === 0
+        ? `Last run: nothing saved.${run.reason ? ` ${run.reason}` : ""}`
+        : `Last run: saved ${run.saved}.${run.reason ? ` ${run.reason}` : ""}`;
     case "skipped":
       return `Last run skipped. ${run.reason ?? ""}`.trim();
     case "failed":
diff --git a/packages/contracts/src/environment.ts b/packages/contracts/src/environment.ts
index 8a0442a5bb538..5317841d7c949 100644
--- a/packages/contracts/src/environment.ts
+++ b/packages/contracts/src/environment.ts
@@ -81,7 +81,7 @@ export const ExecutionEnvironmentCapabilities = Schema.Struct({
   projectKnowledgeHistory: Schema.optionalKey(Schema.Boolean),
   /** projectKnowledge.readDeliveries and the recorded copy delivered with sent messages. */
   projectKnowledgeDeliveries: Schema.optionalKey(Schema.Boolean),
-  /** Assisted proposals from settled runs, the origin list filter and status.assistedLearning. */
+  /** Automatic facts from settled runs, the origin list filter and status.assistedLearning. */
   projectKnowledgeAssistedLearning: Schema.optionalKey(Schema.Boolean),
   projectCode: Schema.optionalKey(Schema.Boolean),
   /** Extended queries are absent from servers that only advertise the original code index. */
diff --git a/packages/contracts/src/projectKnowledge.test.ts b/packages/contracts/src/projectKnowledge.test.ts
index 92d16d03f99c2..f124fd571695d 100644
--- a/packages/contracts/src/projectKnowledge.test.ts
+++ b/packages/contracts/src/projectKnowledge.test.ts
@@ -1,12 +1,22 @@
 import { describe, expect, it } from "vite-plus/test";
 import * as Schema from "effect/Schema";
 
-import { ProjectKnowledgeListResult, ProjectKnowledgeDelivery } from "./projectKnowledge.ts";
+import {
+  ProjectKnowledgeDelivery,
+  ProjectKnowledgeListInput,
+  ProjectKnowledgeListResult,
+  ProjectKnowledgeMutation,
+  ProjectKnowledgeStatus,
+} from "./projectKnowledge.ts";
 
 const decodeDelivery = Schema.decodeUnknownSync(ProjectKnowledgeDelivery);
 
 const decodeList = Schema.decodeUnknownSync(ProjectKnowledgeListResult);
 const encodeList = Schema.encodeSync(ProjectKnowledgeListResult);
+const decodeListInput = Schema.decodeUnknownSync(ProjectKnowledgeListInput);
+const decodeMutation = Schema.decodeUnknownSync(ProjectKnowledgeMutation);
+const decodeStatus = Schema.decodeUnknownSync(ProjectKnowledgeStatus);
+const encodeStatus = Schema.encodeSync(ProjectKnowledgeStatus);
 
 const fact = (origin: Record<string, unknown>) => ({
   id: "fact-1",
@@ -22,7 +32,7 @@ const fact = (origin: Record<string, unknown>) => ({
   entities: [],
   tags: [],
   sources: [],
-  status: "proposed",
+  status: "saved",
   forgottenAt: null,
 });
 
@@ -47,6 +57,63 @@ describe("KnowledgeOrigin", () => {
   });
 });
 
+describe("fact lifecycle", () => {
+  it("encodes and decodes saved directly", () => {
+    const list = { facts: [fact({ kind: "user" })], nextCursor: null };
+    expect(encodeList(decodeList(list)).facts[0]?.status).toBe("saved");
+  });
+
+  it.each(["proposed", "confirmed"])("rejects %s as a current status", (status) => {
+    expect(() =>
+      decodeList({ facts: [{ ...fact({ kind: "user" }), status }], nextCursor: null }),
+    ).toThrow();
+    expect(() => decodeListInput({ projectId: "project-1", status })).toThrow();
+  });
+
+  it("has no confirm mutation", () => {
+    expect(() =>
+      decodeMutation({
+        action: "confirm",
+        commandId: "command-1",
+        projectId: "project-1",
+        factId: "fact-1",
+        expectedRevision: 1,
+      }),
+    ).toThrow();
+  });
+
+  it("reports saved counts and runs without pending counters", () => {
+    const status = {
+      counts: { saved: 2, superseded: 1, forgotten: 0 },
+      trustBands: { low: 0, guarded: 0, steady: 2, high: 0 },
+      recallThreshold: 0.3,
+      belowRecallThreshold: 0,
+      retrievals: 0,
+      feedback: { helpful: 0, unhelpful: 0 },
+      unavailableSources: 0,
+      unavailableSourcesCoverage: "complete",
+      assistedLearning: {
+        lastRun: {
+          threadId: "thread-1",
+          runId: "run-1",
+          runOutcome: "completed",
+          result: "ran",
+          reason: null,
+          saved: 2,
+          at: "2026-10-22T00:00:00.000Z",
+        },
+      },
+    };
+    expect(encodeStatus(decodeStatus(status))).toEqual(status);
+    expect(() =>
+      decodeStatus({
+        ...status,
+        counts: { proposed: 1, confirmed: 1, superseded: 1, forgotten: 0 },
+      }),
+    ).toThrow();
+  });
+});
+
 const oldDelivery = {
   attemptId: "attempt",
   runId: "run",
diff --git a/packages/contracts/src/projectKnowledge.ts b/packages/contracts/src/projectKnowledge.ts
index a56c52c9eb94f..3780a2b47d103 100644
--- a/packages/contracts/src/projectKnowledge.ts
+++ b/packages/contracts/src/projectKnowledge.ts
@@ -47,12 +47,11 @@ const knownKnowledgeOrigins = [
     sessionId: TrimmedNonEmptyString,
     label: Schema.String,
   }),
-  // Written only by the server's assisted learning job, never accepted from a client or agent.
+  // Written only by the server's automatic learning job, never accepted from a client or agent.
   Schema.Struct({
     kind: Schema.Literal("assisted"),
     threadId: ThreadId,
     runId: RunId,
-    // Candidates from a failed or interrupted run are marked so a reviewer weighs them accordingly.
     runOutcome: Schema.Literals(["completed", "failed", "interrupted"]),
   }),
 ] as const;
@@ -98,6 +97,8 @@ export const ProjectFactFields = Schema.Struct({
   kind: ProjectFactKind,
   sources: Schema.Array(KnowledgeFactSource).check(Schema.isMaxLength(16)),
 });
+export const ProjectFactStatus = Schema.Literals(["saved", "forgotten"]);
+export type ProjectFactStatus = typeof ProjectFactStatus.Type;
 const FactIdentity = {
   id: ProjectFactId,
   origin: KnowledgeOrigin,
@@ -116,7 +117,7 @@ const FactIdentity = {
 export const ActiveProjectFact = Schema.Struct({
   ...FactIdentity,
   ...ProjectFactFields.fields,
-  status: Schema.Literals(["proposed", "confirmed"]),
+  status: Schema.Literal("saved"),
   forgottenAt: Schema.Null,
 });
 const ForgottenProjectFact = Schema.Struct({
@@ -141,7 +142,6 @@ const ProjectKnowledgeMutationSchema = Schema.Union([
     ...ProjectFactFields.fields,
   }),
   Schema.Struct({ action: Schema.Literal("edit"), ...ExistingFact, ...ProjectFactFields.fields }),
-  Schema.Struct({ action: Schema.Literal("confirm"), ...ExistingFact }),
   Schema.Struct({ action: Schema.Literal("forget"), ...ExistingFact }),
   Schema.Struct({ action: Schema.Literal("feedback"), ...ExistingFact, helpful: Schema.Boolean }),
   // Creates replacementFactId and links it to factId, which keeps its content (unlike forget).
@@ -179,14 +179,13 @@ export const ProjectKnowledgeMutationResult: Schema.Codec<
 > = ProjectKnowledgeMutationResultSchema;
 export const ProjectKnowledgeListInput = Schema.Struct({
   projectId: ProjectId,
-  status: Schema.optionalKey(Schema.Literals(["confirmed", "proposed", "forgotten"])),
+  status: Schema.optionalKey(ProjectFactStatus),
   afterId: Schema.optionalKey(ProjectFactId),
   limit: Schema.optionalKey(Schema.Int.check(Schema.isBetween({ minimum: 1, maximum: 100 }))),
   includeSuperseded: Schema.optionalKey(Schema.Boolean),
-  /** Only facts whose message citations no longer resolve (archived conversations excluded).
-      Without an explicit status this spans proposed and confirmed facts. */
+  /** Only facts whose message citations no longer resolve. */
   sourceState: Schema.optionalKey(Schema.Literal("unavailable")),
-  /** Only facts proposed by assisted learning. */
+  /** Only facts saved by automatic learning. */
   origin: Schema.optionalKey(Schema.Literal("assisted")),
 });
 export type ProjectKnowledgeListInput = typeof ProjectKnowledgeListInput.Type;
@@ -225,6 +224,7 @@ export const ProjectFactHistoryEvent = Schema.Literals([
   "baseline",
   "created",
   "edited",
+  // Past reviews only; nothing records this any more.
   "confirmed",
   "forgotten",
   "feedback",
@@ -251,7 +251,7 @@ export const ProjectFactHistoryEntry = Schema.Struct({
 export type ProjectFactHistoryEntry = typeof ProjectFactHistoryEntry.Type;
 export const ProjectFactSummary = Schema.Struct({
   id: ProjectFactId,
-  status: Schema.Literals(["proposed", "confirmed", "forgotten"]),
+  status: ProjectFactStatus,
   revision: PositiveInt,
   excerpt: Schema.NullOr(Schema.String.check(Schema.isMaxLength(320))),
 });
@@ -281,20 +281,18 @@ export const AssistedLearningRun = Schema.Struct({
   runOutcome: AssistedLearningRunOutcome,
   result: Schema.Literals(["ran", "skipped", "failed"]),
   reason: Schema.NullOr(Schema.String),
-  proposed: NonNegativeInt,
-  suggested: NonNegativeInt,
+  saved: NonNegativeInt,
   at: IsoDateTime,
 });
 export type AssistedLearningRun = typeof AssistedLearningRun.Type;
 export const ProjectKnowledgeStatus = Schema.Struct({
-  // proposed and confirmed exclude superseded facts, which are counted on their own.
+  // saved excludes superseded facts, which are counted on their own.
   counts: Schema.Struct({
-    proposed: NonNegativeInt,
-    confirmed: NonNegativeInt,
+    saved: NonNegativeInt,
     superseded: NonNegativeInt,
     forgotten: NonNegativeInt,
   }),
-  // Recallable facts (proposed or confirmed, not superseded) by trust quartile.
+  // Recallable facts (saved, not superseded) by trust quartile.
   trustBands: Schema.Struct({
     low: NonNegativeInt,
     guarded: NonNegativeInt,
@@ -309,11 +307,7 @@ export const ProjectKnowledgeStatus = Schema.Struct({
   unavailableSourcesCoverage: Schema.Literals(["complete", "partial"]),
   // Optional so older servers still decode.
   assistedLearning: Schema.optionalKey(
-    Schema.Struct({
-      // Unreviewed assisted proposals, excluding superseded ones.
-      suggested: NonNegativeInt,
-      lastRun: Schema.NullOr(AssistedLearningRun),
-    }),
+    Schema.Struct({ lastRun: Schema.NullOr(AssistedLearningRun) }),
   ),
 });
 export type ProjectKnowledgeStatus = typeof ProjectKnowledgeStatus.Type;
@@ -511,7 +505,6 @@ export class ProjectKnowledgeError extends Schema.TaggedError<ProjectKnowledgeEr
       "forgotten",
       "invalid_source",
       "fact_size",
-      "user_confirmation_required",
       "wrong_environment",
       "wrong_project",
       "stale_context",
@@ -539,8 +532,6 @@ export class ProjectKnowledgeError extends Schema.TaggedError<ProjectKnowledgeEr
         return "Fact text and entities must fit within 16 KiB, and each tag within 256 UTF-8 bytes.";
       case "invalid_source":
         return "The selected source is unavailable or has changed.";
-      case "user_confirmation_required":
-        return "A user must confirm proposed knowledge.";
       case "wrong_project":
         return "This context belongs to another project.";
       case "wrong_environment":
diff --git a/packages/contracts/src/settings.ts b/packages/contracts/src/settings.ts
index 527b4cb25f7c5..e59946e742476 100644
--- a/packages/contracts/src/settings.ts
+++ b/packages/contracts/src/settings.ts
@@ -1351,8 +1351,8 @@ export const ServerSettings = Schema.Struct({
   pstackEnabled: Schema.Boolean.pipe(Schema.withDecodingDefault(Effect.succeed(false))),
   /** Off by default: turning it on moves skill folders out of the user's Claude and Codex homes. */
   skillsLibraryEnabled: Schema.Boolean.pipe(Schema.withDecodingDefault(Effect.succeed(false))),
-  /** Off by default: each settled run in an opted-in project costs a text generation call. */
-  enableAssistedLearning: Schema.Boolean.pipe(Schema.withDecodingDefault(Effect.succeed(false))),
+  /** On by default: each settled run costs a text generation call unless its project opts out. */
+  enableAssistedLearning: Schema.Boolean.pipe(Schema.withDecodingDefault(Effect.succeed(true))),
   /** Off by default: summarizing older history costs text generation calls. */
   enableConversationSummaries: Schema.Boolean.pipe(
     Schema.withDecodingDefault(Effect.succeed(false)),
diff --git a/packages/provider-core/src/server/orchestrationInstructions.ts b/packages/provider-core/src/server/orchestrationInstructions.ts
index 835a39a73688c..5d2161df3288f 100644
--- a/packages/provider-core/src/server/orchestrationInstructions.ts
+++ b/packages/provider-core/src/server/orchestrationInstructions.ts
@@ -3,7 +3,7 @@ import type { ProviderInteractionMode } from "@t3tools/contracts";
 export const T3_CODE_PROJECT_INTELLIGENCE_INSTRUCTIONS = `### Project intelligence
 
 - Use intelligence when the task needs project history or relationships; do not prefetch it for a routine edit. Resolve an unknown project with \`t3_project_list\` and \`t3_project_read\`.
-- Prior decisions, conventions, and constraints: \`t3_knowledge_search\`, then verify message citations with \`t3_knowledge_read_source\`. Recall is on demand. When asked to remember project knowledge, use \`t3_knowledge_mutate\` with exact message citations when available; an agent proposal is not a user confirmation.
+- Prior decisions, conventions, and constraints: \`t3_knowledge_search\`, then verify message citations with \`t3_knowledge_read_source\`. Recall is on demand. When asked to remember project knowledge, use \`t3_knowledge_mutate\` with exact message citations when available. It saves immediately, with no approval step; a saved fact records its origin, not proof that it is true.
 - Symbols, callers, dependencies, affected code, or reaching tests: \`t3_project_code_query\`. Pass the thread id for its registered worktree. Check coverage, freshness, and edge evidence; static candidates are not proof of runtime behavior.
 - For a file's checkpoint history, use \`t3_project_file_history\`. A checkpoint does not establish authorship.
 - A handoff summary is model-written. Use \`t3_conversation_summary_expand\` for the original messages; list or describe summaries to locate one. If a source is unavailable, report that rather than quoting a stale excerpt.