Stop re-explaining your projects
to the AI every session.

Yes — Project Brain is free and open-source under the MIT licence. No sign-up, no payment, no account.

No. It's a Claude Code skill plus a convention: a .project-brain/ folder of plain markdown you can read, edit and commit yourself. There's no Project Brain server and no telemetry — Claude reads the files locally, the same way it reads any file in your project.

A flat doc (CLAUDE.md, notes.md) is loaded in full every session, so it grows until it's both expensive and hard to navigate. Project Brain keeps only a tiny index loaded eagerly and pushes the detail into topic files read on demand. On top of that, status carries the outcome (✓ verified / ✗ failed / ⚠ in-progress) and approaches are versioned instead of overwritten — so it knows what actually worked, not just what was written down.

The opposite, when used as intended. Only the small index is loaded eagerly; topic files are read on demand. An optional one-time "deep backfill" reads a codebase up front — the skill warns you before doing it.

No. Topic files are cold storage — they cost nothing until something opens them, so the per-session weight is bounded by the index, not by how much history you keep. Unlike a flat notes file that gets heavier every session, the brain stays light. To tidy a dead topic you archive it (drop its one line from the index, keep the file). Nothing auto-deletes — your memory is durable by design.

Not really. Claude reads the map at the start of a session, checks it before redoing work, and updates it when a unit of work is done. A bundled brain-check validator catches broken links or status drift if you want to run it — but there's no manual bookkeeping required.

Both. One brain can catalog many projects on a single server, or just a single repo. Multi-project is the default shape — each project is a section in the index.

Yes — it's plain markdown, so it's language-agnostic. Setup detects projects from common signals (package.json, pyproject.toml, go.mod, Cargo.toml, composer.json, git), so it works whatever you build in.

Your call, per project. A real brain ends up holding infra details (DB names, ports, server paths), so you decide whether to commit .project-brain/ — travels with the repo and your team — or keep it out of version control. The skill's own repo ships a .gitignore that ignores .project-brain/, so you never accidentally publish a real brain.

Inside Claude Code. The skill lives in ~/.claude/skills/ per machine; the memory lives in your workspace. It's plain markdown, so your knowledge isn't locked into a proprietary format — built for Claude Code, but yours to keep.