Overview

zero-memory is a shared, persistent memory for LLM coding agents. It is exposed through the Model Context Protocol (MCP), so any MCP-compatible tool (Claude Code, Cursor, Codex CLI, VS Code, and others) reads and writes the same memory instead of rebuilding context from scratch in every session.

Why

Every new chat session starts from zero. The decision you explained yesterday, the gotcha that cost an afternoon, the convention nobody wrote down: each session re-derives them, usually differently.

Writing them down is the easy half. The hard half is that they have to come back on their own — in whatever client opens next, to whoever is working, and only to the people they belong to.

Key ideas

Memory comes in two tiers. Private fragments are visible only to their author. Shared fragments carry provenance: who wrote them, when, and from which conversation. Sharing is always an explicit action.

Scope membership is explicit too. Alongside private personal memory there are owner-namespaced project and team scopes that can be shared. A matching project name never grants access, the session's project is the write default, and a write with no resolvable target is refused rather than mis-filed. The details are in Scopes and isolation.

There are several ways in: deliberate agent writes, terminal quick capture, deterministic import, repository bootstrap, and opt-in transcript ingest. That last one is disabled by default.

Everything is MCP-first. Memory, briefing, hygiene, rules, import/export, and metrics tools all work in any MCP-compatible client — see the memory model. Standing rules are served over the network: promoted rules use native MCP instructions where the client can carry them, and build_context everywhere else. Hooks add delivery timing and offline resilience.

The backend is Postgres via Supabase. Recall runs a hybrid RAG pipeline over pgvector and full-text search, on top of an entity graph as edge tables, Row Level Security for scoping, and Realtime for live team updates.

How the docs are organized

• Getting Started — connect a client to a running instance, or deploy your own stack.
• Concepts — the memory model and the design decisions behind it.
• Guides — day-to-day workflows: briefings, import/export, session receipts.
• Dashboard — the web UI for browsing and curating memory.
• Memory Hygiene — how the store stays clean over time.
• Platform — architecture, observability, testing, and internals.