you

A fictional, open-source memory dataset for Previously. One anonymous life, told as dated time slices from 2024 to 2026.

What this is

you is the demo memory corpus behind the Previously playground. It follows an unnamed protagonist — addressed simply as you — through roughly two and a half years of work, running, relationships, family, and everyday life. The story is optimistic at its core, but it includes real setbacks — injury, self-doubt, conflict, a family health scare — because a memory that only stores happy moments is not a real memory.

The data is entirely synthetic. No real person, city, company, or private information is represented here.

The dataset is periodically updated: new slices are appended as the story moves forward, so returning visitors always have something new to recall.

Repository layout

story/story.yaml            the story bible: persona, strands, events, real-world anchors
reference/persona.md        canonical persona reference for downstream repos (site copy, docs, demos)
manifest.json               machine-readable index of personas, topics, and the slice tree
user/
  user/profile.md           protagonist profile
  episodic/
    timeline.md             every slice on one page, newest first
    strands.json            tag → slice-count index
    current-previously.md   the memory snapshot as of the latest slice
    slices/YYYY/MM/DD/HHMM/
      previously.md         memory snapshot (populated on the latest slice only, by design)
      timeline/core.md      the conversation itself: frontmatter + dated turns
      timeline/agent.md     agent-side timeline (reserved, currently empty)
  evolution/
    direction.md            the evolution direction: a user portrait over six
                            fixed dimensions plus a pool of falsifiable
                            trait-level hypotheses (v1.0)
  agent-playbooks/
    recall.md               evolved working notes for the recall sub-agent
    search.md               evolved working notes for the search sub-agent
    thinkdeep.md            evolved working notes for the think-deep sub-agent
  .loom-state.json          generator resume state, kept so updates can continue incrementally

How it is made

The dataset is generated by Loom, Previously's open dataset generator:

loom generate --story story/story.yaml --out user --all --format previously --resume

Loom is a modern re-implementation of the methodology behind LoCoMo (ACL 2024), the de-facto standard benchmark for very long-term conversational memory: a persona plus a temporal event graph drives multi-session generation, while all dates, weekdays, and timestamps are computed in code — never written by the model. Loom writes Previously's native slice format (used here), and can also emit LoCoMo-compatible JSON, so anyone evaluating long-term memory systems can generate their own datasets with it.

Story strands

1. Work / Side project — a small tool born from a personal frustration, from CSV script to launched product.
2. Running / Health — from winded on the stairs to finishing a half-marathon.
3. Alex — a close relationship, with its own friction and growth.
4. Family — weekly calls with Mom, a health scare, and learning to show up.
5. Friends / Community — Priya, Marcus, and the people who shape the weeks.

Real-world anchors

The story weaves in real events as shared cultural reference points:

• Paris 2024 Summer Olympics
• DeepSeek-R1 release (January 2025)
• Milano-Cortina 2026 Winter Olympics
• FIFA World Cup 2026

Dates are fixed in the story bible; the model is only asked to have the characters talk about them.

Usage

you is consumed by the Previously demo service, where playground presets map to specific recall queries against this corpus. If you build something on top of this dataset, start with reference/persona.md — it is the canonical description of who the protagonist is and what actually happens, and all derived copy should stay consistent with it.

License

MIT © Previously Lab