Stop re-explaining your project to every new agent

What it does

The toolkit drops a BRAIN.md protocol file and a brain/ directory into any project. A zero-dependency Node CLI then handles all reads and writes, storing durable knowledge—architectural decisions, requirements, constraints—as plain Markdown in the repo. Any agent that can read files (Claude Code, Codex, Cursor, and others) learns the protocol from BRAIN.md and picks up where the last session left off.

The interesting bit

The design is “correct by construction”: because the CLI is the sole writer, frontmatter is always generated and the update-truth command atomically rewrites the current understanding while appending a timeline entry. This structurally prevents malformed edits or silent changes without a trace, so there is no need for a separate validator. It treats the chat log as ephemeral and the repo as the single source of truth.

Key highlights

• Agent-agnostic contract: any agent that reads BRAIN.md can use the brain, no special integration required.
• Append-only timeline plus rewritable compiled truth per page, so context evolves but never erases its own history.
• Pure Markdown stored in git; no database, no server, and the brain travels with the codebase.
• Location-independent: a brainRoot setting can redirect the brain directory without confusing tooling.
• Includes opt-in skills for bootstrapping knowledge from existing code/docs and ingesting conversations.

Caveats

• The project is early (v0.2.0) and some runtime integrations are noted as “more to come.”
• The brain starts empty after init; seeding it with existing knowledge is a deliberate, separate step.
• Stewardship and maintenance belong to MindMux, despite the spec’s neutral naming and open-source Apache-2.0 license.

Verdict

Worth a look if you juggle multiple agents or long-running projects and are tired of repeating the same constraints in every new chat. Skip it if you already have a wiki or documentation system your agents reliably read and update.