From 3a8d49389dc66cdc0fcb910a6c43f153ea95f51b Mon Sep 17 00:00:00 2001 From: Grivn Date: Tue, 29 Sep 2026 04:59:51 +0800 Subject: [PATCH] docs: add an experimental memory-agent entry Add experimental/ for research previews that are not part of the mnemon binary or its releases, with a first entry, memory-agent/, in English and Chinese. It describes Mnemon as a memory agent on DeepSeek Harness, with Jev as the System 1 decision model and an LLM as System 2, and links to the paper, code and run records in Grivn/mnemon-memory-agent. The root README and its Chinese counterpart list it under Documentation. --- README.md | 1 + docs/zh/README.md | 1 + experimental/README.md | 8 ++++++ experimental/memory-agent/README.md | 35 +++++++++++++++++++++++ experimental/memory-agent/README.zh-CN.md | 27 +++++++++++++++++ 5 files changed, 72 insertions(+) create mode 100644 experimental/README.md create mode 100644 experimental/memory-agent/README.md create mode 100644 experimental/memory-agent/README.zh-CN.md diff --git a/README.md b/README.md index 052d21de..dd292d51 100644 --- a/README.md +++ b/README.md @@ -535,6 +535,7 @@ See [Development and Deployment](docs/DEPLOYMENT.md) for Docker, Compose, Ollama - [Memory Usage & Reference](docs/USAGE.md) — root Memory commands, import, receipts, and embedding support - [Memory Import Guide](docs/IMPORT.md) — schema and LLM prompt for importing historical chats - [Architecture Diagrams](docs/diagrams/) — system architecture, pipelines, lifecycle management +- [Memory Agent (experimental)](experimental/memory-agent/) — research preview of Mnemon as a memory agent on DSH, with Jev as its System 1 decision model; paper, code and run records ## Star History diff --git a/docs/zh/README.md b/docs/zh/README.md index a42b0a03..ff26ac1e 100644 --- a/docs/zh/README.md +++ b/docs/zh/README.md @@ -462,6 +462,7 @@ make help # 显示所有目标 - [Memory 用法与参考](USAGE.md) — 根级 Memory 命令、导入、回执与嵌入向量支持 - [记忆导入指南](IMPORT.md) — 导入历史聊天的 schema 与 LLM 提取提示词 - [架构图](../diagrams/) — 系统架构、记忆/召回流程、四图模型、生命周期管理 +- [Memory Agent（实验性）](../../experimental/memory-agent/README.zh-CN.md) — Mnemon 以记忆 Agent 形态运行在 DSH 上的研究预览，由 System 1 决策模型 Jev 做判断；含论文、代码与运行记录 ## 参考文献 diff --git a/experimental/README.md b/experimental/README.md new file mode 100644 index 00000000..575e2374 --- /dev/null +++ b/experimental/README.md @@ -0,0 +1,8 @@ +# Experimental + +Research previews built around Mnemon. They are not part of the supported `mnemon` binary or its releases, and they +may change or move without notice. + +| Directory | What it is | +|---|---| +| [`memory-agent/`](memory-agent/) | Mnemon as a memory agent on DeepSeek Harness: a fast decision model (System 1) judges raw records, and an LLM (System 2) plans and answers | diff --git a/experimental/memory-agent/README.md b/experimental/memory-agent/README.md new file mode 100644 index 00000000..8e61ae2f --- /dev/null +++ b/experimental/memory-agent/README.md @@ -0,0 +1,35 @@ +# Memory Agent (experimental) + +

English · 中文

+ +A research preview of Mnemon as a **memory agent**. It keeps conversations as raw, dated records and divides the work +of memory the way dual-process accounts divide thinking: + +- **System 1**: a fast decision model, [Jev](https://typesafe.ai/blog/introducing-system-one-models-and-jev) from + TypeSafe AI, makes many small yes/no judgments about the records a search returns, such as whether a record is + needed or no longer current. +- **System 2**: an LLM writes a few search queries, names what the reply needs and composes the answer. +- **Consolidation**: a background pass consolidates each record once into an index linked to the records. + +The agent runs as a second [DeepSeek Harness](https://github.com/deepseek-ai/deepseek-harness) (DSH) instance beside +the main agent and hands it a small View at every turn. + +**Code, paper and run records:** [Grivn/mnemon-memory-agent](https://github.com/Grivn/mnemon-memory-agent) + +## Results + +- **LoCoMo 91.7%**, first of the 15 systems that answer with gpt-4.1-mini (Mnemon and the 14 systems OmniMemEval + re-evaluated), from 3.8k tokens of context per question. +- **LongMemEval-S 94.4%** with DeepSeek-V4.1-Flash answering, on par with the best published results. +- On the same records, Jev separates gold evidence better than DeepSeek and gpt-4.1-mini and is 3–11× faster. + +Details, baselines and caveats are in the paper, *Mnemon: Raw Records, Fast Judgments, Slow Thoughts* +([PDF](https://github.com/Grivn/mnemon-memory-agent/blob/master/docs/paper/main.pdf); arXiv link to follow). + +## Status + +- A frozen research snapshot. It is not part of the `mnemon` binary, its releases or Mnemon Agency, and the evaluated + system does not use the `mnemon` binary. +- Its results will be brought step by step into Mnemon and [dsh-mnemon](https://github.com/omdsh-dev/dsh-mnemon). +- Next, DSH will serve as a micro agent kernel, with Mnemon running on it as a memory agent, the form this preview + already takes. diff --git a/experimental/memory-agent/README.zh-CN.md b/experimental/memory-agent/README.zh-CN.md new file mode 100644 index 00000000..669d4340 --- /dev/null +++ b/experimental/memory-agent/README.zh-CN.md @@ -0,0 +1,27 @@ +# Memory Agent（实验性） + +

English · 中文

+ +这是 Mnemon 以**记忆 Agent** 形态运行的研究预览。它把对话保存为带日期的原始记录，并按"双系统"理论的方式分工： + +- **System 1**：快速的决策模型 [Jev](https://typesafe.ai/blog/introducing-system-one-models-and-jev)（TypeSafe AI），对检索返回的记录做大量简单的是非判断，例如这条记录是否需要、是否已经不再成立。 +- **System 2**：一个 LLM，负责写少量检索查询、说明回复需要什么，并组织最终回答。 +- **后台整合**：把每条记录整合一次，建成链接回记录的索引。 + +它作为第二个 [DeepSeek Harness](https://github.com/deepseek-ai/deepseek-harness)（DSH）实例运行在主 Agent 旁边，每一轮交给主 Agent 一个精简的 View。 + +**代码、论文与运行记录：**[Grivn/mnemon-memory-agent](https://github.com/Grivn/mnemon-memory-agent) + +## 结果 + +- **LoCoMo 91.7%**：在用 gpt-4.1-mini 答题的 15 个系统（Mnemon 与 OmniMemEval 重测的 14 个系统）中排名第一，每题只用 3.8k tokens 上下文。 +- **LongMemEval-S 94.4%**：由 DeepSeek-V4.1-Flash 答题，与已公开的最好成绩持平。 +- 在同一批记录上，Jev 区分关键证据的效果优于 DeepSeek 和 gpt-4.1-mini，速度快 3–11 倍。 + +细节、对比基线与注意事项见论文《Mnemon: Raw Records, Fast Judgments, Slow Thoughts》（[PDF](https://github.com/Grivn/mnemon-memory-agent/blob/master/docs/paper/main.pdf)；arXiv 链接稍后补充）。 + +## 状态 + +- 这是冻结的研究快照，不属于 `mnemon` 程序、它的发布版本或 Mnemon Agency；评测的系统没有用到 `mnemon` 二进制。 +- 研究成果将逐步输送到 Mnemon 和 [dsh-mnemon](https://github.com/omdsh-dev/dsh-mnemon)。 +- 下一步，DSH 将作为微型 Agent 内核，Mnemon 作为记忆 Agent 运行在它之上；这个预览就是这种形态。