Pattern Automation
← Blog

Designing agent memory for multiplayer on Neuro OS

Single-user memory breaks in group rooms — per-person notes, front-of-card vs drawer, and direct file recall instead of similarity search alone.

Discuss this post in AI

Send a pre-filled prompt to ChatGPT, Claude, Gemini, or Perplexity — get a summary, ask follow-ups, or compare ideas from this guide.

The wrong question

Almost every AI memory system answers: “What do I know?”

Neuro OS group work needs: “What do I know about this person, right now?”

That sounds like a small shift. It is not. Most memory designs — long-term note plus recent log, or retrieval over everything — assume one human and one assistant. Put the same notebook in a group room and the implicit subject vanishes.

A notebook for one

The obvious first design: durable facts note + running daily log. Agent reads the note at session start.

For 1:1, it works. It is a personal notebook — everything is implicitly about you.

In a shared task thread with multiple humans and roles:

  • “Prefers short replies, no emoji” — whose preference?
  • Two people named Mike — search returns text, not a person
  • “Mike is planning Tokyo” stays findable long after the trip ended

The fix is not a bigger notebook. Memory must know who each fact belongs to.

Why “search everything” does not rescue it

Retrieval-first memory — save all, embed, fetch similar chunks — powers impressive single-user systems. It matches text to text.

In shared space you need who, and is it still true?

Consider two different people named Mike in the same company thread. Search for “Mike” ranks passages by similarity. The search can be flawless; the answer still wrong — because what came back was language, not identity.

Retrieval also ages quietly: stale facts stay equally findable.

We could patch search to be identity-aware. On Neuro OS we changed what memory is.

Memory you can browse

Per-person files — ordinary notes the role opens directly. No hidden index. Identity is whose folder this is, not a guess recovered from embeddings.

Within each person’s memory, one judgment per fact:

  • Front of card — matters every conversation (tone, serious allergy, timezone)
  • Drawer — real, retrievable, but not clipped to every turn

A good front desk keeps the card front short. A memory that surfaces everything surfaces nothing useful.

Opening the drawer

Classic file-memory failure: agent decides the drawer is irrelevant and never looks.

Coding agents solved a parallel problem: they do not build an elaborate index for one function — they grep the repo. Plain text, exact, in the moment.

Neuro OS roles recall the same way when a specific question arises: open the person’s drawer files and read. Not “search everything” under a new name — a place per person and a direct look when needed.

Does it work?

Early live tests on Neuro OS showed:

  • Drawer facts (allergy, pet name, travel plan) recalled when asked — agent looked instead of guessing
  • Single message with mixed details filed sensibly: communication preference on front, time-bound project in drawer
  • Legacy single-notebook memory reorganized into per-person shape during ordinary work

Caveats: strong-model conversations, not thousands of measured runs. Open questions:

  • Do smaller models reliably open the drawer?
  • Who curates what stays on the card front — agent alone or human policy?

Design for teams on Neuro OS

Layer Single-user habit Multiplayer fix
Unit one notebook folder per person
Recall similarity rank identity + direct read
Freshness implicit time-bound facts in drawer
Governance user edits note git-backed skills + role policy

Role memory composes with shared task state — public progress in the thread; private per-person notes for how to talk to each member.

The larger point

We did not move away from search because search was bad. The unit was wrong. Shared rooms make “who, and is it still true” the first question.

Plain notes, one set per person, read and searched directly — the same move engineers trust in codebases.

Related: Agent collaboration protocol · The group workspace · Personal AI agent

Explore Neuro OS →

More from Blog