It does not index your code — it records what you touched together and makes those paths cheap to reach, which is why the git-history bootstrap matters: a cold network knows nothing. The system-wide daemon that watches other editors is deliberately not installed for you, because it registers a LaunchAgent and monitors your project directories.
A memory layer for coding agents built on Hebbian learning: files, tools and errors are neurons, accessing them together forms synapses, frequently used paths strengthen, and unused connections decay.
- 6 MCP tools — record, recall, error, predict_next, stats and decay
- Passive learning through Claude Code hooks: a PostToolUse hook learns from every file read, edit and search, a UserPromptSubmit hook injects recall into prompts
- An error-fix memory — which files fixed which error, offered again the next time a similar one appears
- Tool-sequence prediction: after enough Grep-Read-Edit chains it predicts what comes next and pre-loads the likely files
- The same operations on the command line — `brainbox recall`, `brainbox record`, `brainbox stats`, `brainbox hubs`, `brainbox highways`, `brainbox graph`
- `brainbox bootstrap --repo /path/to/project --imports` seeds the network from git co-changes and the import graph, so it doesn't start from nothing
`npm install brainbox-hebbian`. The postinstall step registers the MCP server, adds both Claude Code hooks and creates `~/.brainbox/`. For other agents, run the server standalone with `npx tsx node_modules/brainbox-hebbian/src/mcp.ts`. The macOS file-watching daemon is separate and opt-in via `brainbox daemon install`; `brainbox uninstall` removes the hooks and server while keeping the database.
One command — npm install brainbox-hebbian