BRAIN.md

---

This repository is the toolkit, not a brain itself. Install it once, then in any project run brain init: it scaffolds a BRAIN.md protocol file, a brain/ directory, and default-wires agent config files into your repo. From then on, any coding agent — Claude Code, Codex, anything that reads files — learns to use that brain just by reading the project's BRAIN.md. The brain is plain Markdown, lives in the repo, and outlives every session.

Why a brain

A coding agent's knowledge lives nowhere durable. The reasons behind a decision, the constraints you agreed on, the path not taken — they sit in chat logs and in your head, and they vanish the moment the session ends. The next agent starts from zero.

A brain fixes that. It is the project's persistent memory: the durable decisions, requirements, and constraints, written down as plain Markdown next to the code.

• Repo-native — Markdown that lives in your project and travels in git, with or without a runtime on top.
• Agent-agnostic — the contract is a file (BRAIN.md). Any agent that can read it can use the brain.
• Correct by construction — every write goes through the brain CLI, so the brain's invariants can't be broken by a malformed edit. There is no validator because none is needed.

The test for what belongs in it: will this still matter in six months, and is it hard to reconstruct from the code itself? If yes, it goes in the brain. Pure implementation details and anything readable straight from the code and git history stay where they are.

Quick start

1. Install the tools once (global) — no clone required. This puts brain on your PATH and copies skills into every detected agent (~/.claude/skills, …):

npm install -g @mindmux/brain-md
brain setup -y
# reverse: brain uninstall   # never touches any project's brain data

Prefer not to install globally? Use npx for both steps (npx does not leave brain on your PATH):

npx @mindmux/brain-md setup -y
npx @mindmux/brain-md init          # same as step 2 — not bare `brain init`
# reverse: npx @mindmux/brain-md uninstall

From a git checkout of this repo you can still run ./setup (same installer; use --symlink while developing the toolkit).

2. Initialize a project (from the project root; requires the global install from step 1, or use npx @mindmux/brain-md init above):

brain init
# or, after setup only (no global bin): node ~/.claude/skills/brain-page/bin/brain.mjs init

This ensures BRAIN.md, scaffolds empty brain data (brainRoot-aware), and default-wires CLAUDE.md + AGENTS.md (creates them if missing; if they already exist, only updates the marked brain block — never whole-file overwrite). Optionally use the brain-setup skill for the same flow plus a pre-commit hook, or brain install-hooks for a project-local Claude Code SessionStart snapshot (--agent codex for Codex).

3. Seed real knowledge — run the brain-bootstrap skill. On an existing project it reads the code, docs, and git log to draft the root pages and capture key decisions; on a near-empty one it interviews you. (Init leaves content seeding as a deliberate next step.)

4. Work as usual. Maintain the brain while coding: load relevant pages at task start; capture decisions/constraints when they settle; skip pure implementation noise; reverse when overturning. All reads/writes go through the brain CLI following BRAIN.md — never hand-edit brain files.

See it work

The point of a brain is what happens across sessions — a decision made today is still there, with its rationale, weeks later:

You    Let's store config as Markdown, not SQLite — easier to diff and migrate.

Agent  Capturing that as a decision so it outlives this session.
       $ brain create-page --id config-as-markdown --category decision \
             --title "Store config as Markdown, not SQLite"
       ✓ page created · indexed

— three weeks and a fresh session later —

You    Why aren't we using a database for config?

Agent  $ brain read-page config-as-markdown
       We chose Markdown for diff-ability and zero migrations. Here's the original call
       and the trade-offs we weighed …

The brain CLI

Reading and writing the brain both go through one zero-dependency Node CLI (run with node):

brain() { node skills/brain-page/bin/brain.mjs "$@"; }   # or use the global `brain` bin after npm install -g

brain init                                   # BRAIN.md + skeleton + default wire CLAUDE.md / AGENTS.md
brain wire                                   # same default wire (no --agent needed)
brain install-hooks                          # opt-in Claude Code SessionStart snapshot (project-local)
brain uninstall-hooks                        # remove that SessionStart hook
brain install-hooks --agent codex            # opt-in Codex startup/resume/compaction snapshot
brain uninstall-hooks --agent codex          # remove only the Codex hook
brain brain-dir                              # where is the brain?
brain list-pages                             # list pages
brain read-page my-decision                  # read a page
brain create-page --id my-decision --category decision --title "Use X over Y" \
  --summary "why this page exists"
echo "the new understanding" | brain update-truth --id my-decision --summary "why it changed"
brain append-timeline --id my-decision --kind evidence --summary "benchmark confirmed it"
printf '# System architecture\n\n## Overview\n\nRequests enter through the gateway.\n' | brain update-root architecture
brain reindex && brain lint-links

以上為 README 的開頭部分。其餘內容請閱讀原文。 檢視完整 README ↗