Shared, persistent memory for AI coding agents. One memory server speaks MCP, so Claude Code, Cursor, Codex CLI, VS Code, and any other MCP-compatible client read and write the same store instead of rebuilding context from scratch every session.
you ──▶ Claude Code ─┐ ┌──────────────────────────────┐
you ──▶ Cursor ──────┼──▶ zero-memory (MCP) ──┤ pgvector + FTS, RRF-fused │
teammate ─▶ Codex ───┘ write ──▶ │ decisions · gotchas · rules │
◀── recall └─ Postgres · RLS ─────────────┘
Every new session starts from zero. The decision you explained yesterday, the gotcha that cost an afternoon, the convention nobody wrote down — all of it has to be retyped, or it is silently re-derived, usually differently.
Writing it down is the easy half. The hard half is that it has to come back on its own, in whatever tool the next session happens to open, to whoever is working — and only to the people it belongs to.
It remembers decisions with their reasoning, not just facts, alongside gotchas, conventions, preferences, and open loops that stay visible until someone closes them.
A session gets briefed before it starts. build_context returns the standing
rules, the open work, and the decisions behind the code you are about to
touch; hook-capable clients inject it automatically.
Scopes stay honest. Personal, project, and team scopes are owner-namespaced, and access comes from explicit membership, so two projects with the same name never share anything. Row-level security enforces that in Postgres, not in application code.
Sharing is deliberate. A fragment is private to its author until it is shared, and it travels with its provenance: who wrote it, when, and from which conversation. A corrected memory retires the version it replaces, reversibly, so the store converges instead of accumulating contradictions. Content is canonicalized into one internal language, so a memory written in one language is retrievable from another.
The boring parts run themselves: hygiene passes, duplicate detection, decay and reinforcement, conflict surfacing.
Core memory (recall, remember, search, scopes, briefings) runs locally and needs no model-provider key. Model-backed extras (repository bootstrap, hygiene judgements, translation, rule distillation, optional transcript ingest) need a provider credential, and transcript ingest is off by default: a fresh install reads memory and sends nothing.
Membership is granted per person, per role, and the scope id carries its owner:
When two memories disagree and automated hygiene cannot settle it, the pair is surfaced side by side with the judge's confidence and rationale, and the decision stays yours:
Work that outlives a handover gets a card. A project's board holds each piece of work for as long as it runs — its document, the memories and links it points at, and a history in which every move carries its reason. Agents move cards through MCP; the dashboard shows the board and deliberately cannot move anything:
Using an instance someone already runs. You need an account, the client bundle for your platform, and one installer script. No Docker, no checkout: see Install the client.
Running your own instance. Docker for the Supabase backend, a host for
the two application containers, and a TLS edge if it will be reachable beyond
a trusted network: see Deployment.
One invariant matters more than the rest. ZM_PUBLIC_URL must equal the
external URL of the MCP server, because it is the OAuth issuer embedded in
every challenge.
Hacking on the code. Bun ≥ 1.3.11 and Docker:
bun install
bun run devThat starts the local Supabase stack, writes the .env files, provisions a
development user, and brings up the MCP server on :8787, the dashboard on
:3100, and the documentation site on :3200.
A single MCP server exposes memory, briefing, hygiene, rules, import/export,
and metrics tools to every client. Standing rules promoted by an owner travel
through native MCP instructions where the client can carry them, and through
build_context everywhere else. Supabase provides Postgres, auth, and
realtime updates, self-hosted from the compose stack in this repository or on
Supabase cloud.
Retrieval is a hybrid RAG pipeline over atomic memories rather than document chunks: pgvector similarity and Postgres full-text search run as two legs and are combined by reciprocal rank fusion, so an exact identifier still ranks when semantic similarity is weak. Long memories are covered by as many embedding windows as their length needs, because a single vector silently truncates past the model's window. A hit comes back as an excerpt cut from the window that matched, not the whole record. Above it sits an entity graph stored as edge tables.
You can check whether it is actually helping: recall hit rate, what only memory knew, and how the corpus ages are all reported from usage events.
The stack is Turborepo + TypeScript + Next.js, arranged as a DDD/hexagonal workspace:
| Path | What lives there |
|---|---|
apps/server |
the MCP server and its HTTP surface |
apps/web |
the dashboard |
apps/docs |
the documentation site |
apps/watcher |
the client-side binary: hooks, ingest, briefings |
packages/* |
domain, contracts, persistence, retrieval, hygiene, adapters |
supabase/migrations |
schema and row-level-security policies |
infra/ |
development and production compose stacks |
tests/e2e |
the Playwright suite and its isolated stack |
Full documentation lives in docs/ as Markdown and is served as a
site by apps/docs. Start with
the overview, or go straight to
getting started. Worth reading early:
scopes and isolation, the model
that decides who can see what.
Pre-1.0, running in production for its own team, and used daily to build itself. Interfaces still move: the schema evolves through forward-only migrations, and the memory tools are stable in shape but not frozen.
Bug reports, deployment reports from hosts we have never touched, and client adapters are the most useful contributions. Start with CONTRIBUTING.md; contributors sign a CLA on their first pull request, and the Code of Conduct applies everywhere.
Found a security problem? Do not open an issue. Follow SECURITY.md.
Apache License 2.0 — Copyright 2026 Kiwagu.