The memory layer, open source, in one command

Added

• One-command setup. mycobrain-setup memory finds or starts a local Postgres with pgvector, or takes a Supabase, Neon, RDS or other Postgres connection string by reference so it never lands in your shell history. It shows every change and waits for a yes, backs up each file it edits, writes no secret into any of them, and finishes by writing one fact and recalling it. mycobrain-setup memory --uninstall removes exactly what it added.
• Starter templates. Agency, design studio, architecture or landscape firm, solo consultant, or custom, a short guided interview that writes a template for your own business.
• Automatic recall and capture. Claude Code and Codex get hooks: recall at session start and on each prompt, and capture of finished sessions. Cursor, Windsurf and Claude Desktop get an MCP server with two tools, brain_memory_context and brain_memory_write.
• Facts with a source, never overwritten. Every fact records where it came from. A changed fact supersedes the old one, and mycobrain-memory history shows both and the reason.
• Client boundaries. Houses keep clients apart, and only the owner creates them. mycobrain-memory binding add <name> --house <slug> scopes an agent to one client: it cannot read or write any other client's facts. Each runtime gets its own credential, so you can see which tool wrote a fact and revoke one tool without touching the others.
• A background worker that needs no API keys. It reads finished sessions for durable facts using your own Claude or Codex subscription, embeds new facts locally with Ollama, and delivers writes saved while the database was unreachable. Setup can install it as a launchd agent (macOS) or a systemd user service (Linux).
• Dream, off by default and review-first. Dream proposes merges of duplicates, per-client dossiers and cross-client patterns. Nothing changes until you approve it with mycobrain-memory review approve.
• A doctor for memory. mycobrain-doctor --memory checks the config, the database, each runtime's credential, the hooks, the worker and Ollama, with the fix for anything that is off.

Changed

• The MCP server picks its mode from its environment: 15 MCP tools with the engine, 2 in memory-only mode. In a memory-only install its instructions point agents at the memory tools, and an engine tool called anyway answers with a pointer to them.
• The document engine is presented as optional: add it with one docker compose up when you want it.

Proven on a clean machine

Before release we ran a stranger test: a fresh temporary home, a fresh Postgres, the packed npm tarball, and the documented commands, set up as a fictional landscape architecture firm. A decision made in the first session came back unprompted in the third. A change superseded the old fact with its history and reason kept. An agent scoped to one client found nothing from another. With the database stopped, sessions still started and a write was queued and delivered later. A Supabase-style setup with a non-superuser login finished green (on a cluster shared with another memory database, after the one grant the quickstart documents). Uninstall put the user's config files back byte for byte.

Existing engine tool contracts are unchanged. Full details in the repository CHANGELOG.