Marrow

The product context layer for coding agents

Your coding agent has never been in the room.

The why lives in standups, interviews and whiteboards. None of it reaches the agent, so it guesses, and ships what nobody agreed to. Marrow turns the room into your product's source of truth, goals and features traced to where they were decided, keeps it current as the product moves, and puts the agent in it.

Run from source Read the docs

Open source, Apache 2.0 · Runs on one Postgres · Default Claude

Exhibit A · The raw room

EVIDENCE ev_3f9a · interviews/design-partner.md · excerpt · 3 speakers · Append only

  1. Dana:Tell me about the last time the product scared you.
  2. Partner:A teammate hard-deleted our staging project last quarter. One click, confirmed, and it was gone.
  3. Dana:There was no way back?
  4. Partner:Nothing. It purged on confirm. Support could not recover it and we lost a week rebuilding.
  5. Partner:A hard delete on the wrong project is a support fire. Give us a window to undo it.
  6. Dana:Then that's it. Soft delete, 30 days, then purge.
  7. Sam:Agreed. Recoverable for a month, then it is really gone.
  8. Dana:Does 30 days cover your audit window?
  9. Partner:More than enough. A Friday-afternoon mistake just has to be survivable.
  10. Sam:And what purge means for backups still needs its own call. Not today.
  11. Dana:Noted. Park it as open.

A reenactment of the loop, running on this page. Nothing is saved.

Exhibit B · What is still open

EVIDENCE ev_77c1 · standups/2026-06-02.md · excerpt · Append only

  1. Priya:Soft delete shipped to staging, recoverable for 30 days like we agreed. QA restored a deleted project this morning in two clicks.
  2. Priya:The old rule where a delete needed a founder email to reverse is dead, soft delete replaced it.
  3. Marco:The open one is dashboard sessions. I want a 12 hour idle timeout, half our users sit on shared screens in open offices and walk away.
  4. Priya:Keep sessions until they sign out. People live in the dashboard all day and a re-login mid-task kills the flow.
  5. Marco:We did not settle it. Parking it for the security review.

Exhibit C · What the agent reads

The agent starts with a decision gate. It gets the slice for the task, with status and citation on every fact, then checks its diff before review. That is the difference between context and a dump.

MCP · prepare_task({ task, check }) → compact brief

Safe to build decided rows are buildable
Ask human first open rows stop the task
Open Deletes are soft for 30 days, then purged 0.60 Model · Waiting on you, Exhibit A
Decided Pricing is per workspace, flat 1.00 Human ev_9b2e [62–103]
Open Dashboard sessions: 12h idle or persist 0.60 Model
Entity The editor 0.85 Model ev_77c1 [686–706]

Agent: prepare_task says dashboard sessions are still open (ev_77c1). Asking before building.

The slice, not the brain. The agent sees what is decided now and what is still open, and Marrow keeps it current, raising a question when code drifts from a decided goal instead of letting the two quietly diverge. Check mode returns the receipt, the catch event and the accept or dismiss command. Run the same loop →

Put your room on the record.

From a clean machine to a decided fact, no API key for the demo. The only thing you bring is Postgres.

INTAKE · from source

Apache 2.0 One Postgres Default Claude
Star on GitHub Run from source

Support · questions, answered

Still have a question?

The honest answers first. Then the docs, an issue, or a human. We answer.

Does Marrow read my codebase?

No. The room is the input, never the repo. Marrow watches the gap between what the room decided and what the code does, and raises it as a question. The code-memory tools distill from git; we never treat the code as truth.

Does Marrow hold our product goals?

Yes. Product goals and user goals live in the brain as first-class truth, each tied to the feature it serves and traced to where it was decided. Your team writes them, distillation proposes them from the room, and when code drifts from a decided goal Marrow raises a question. It is the source of truth your agent reads before it builds.

Where does my data go?

Into your Postgres, and nowhere else by default. Connector tokens are encrypted at rest. No model provider trains on your data, and you can bring your own keys.

What can it pull from automatically?

Slack, Microsoft Teams, Gmail, Linear, Jira, GitHub Issues, Notion, Figma, Zoom, granola, otter and intercom. Each connector only ever appends evidence, never edits it. You watch the sync state and the cost in the console.

Will it slow my agent down?

No. Retrieval is task-scoped and bounded, fast enough to sit inside the agent loop. Every distill, search and drift check is recorded, so you can see the latency and the cost yourself.

Is the evidence really immutable?

Yes. Raw evidence is append only. There is no update or delete path, in the code or the schema. A correction is a new row, so every decided fact still traces to the exact line it came from.

Can I self-host?

Yes. Apache 2.0, one Postgres, one model key, no extra infra. The managed cloud is for teams who would rather not run it, the same split as temporal and supabase.

Replies in < 1 day Apache 2.0
Open an issue Ask in discussions Email support

The others remember the code. The code was never in the room.1

1. Lore, Tenet and the rest distill from git. The room is the one source they cannot reach.

This page holds 7 facts: 1 Decided · 5 Open · 1 Superseded. Every one traced.

Set in Fraunces, Geist and Geist Mono, self-hosted. Serif is decided, sans is open, mono is evidence. One static page, no build step, no tracking.

— End of file —