RoBrain plugin for Claude Code

Claude's memory keeps your best notes. RoBrain keeps the things you decided not to do — and hands them back right before they're about to happen again.

The plugin wires three lifecycle hooks into Claude Code, all backed by your self-hosted RoBrain stack:

Hook	What it does
SessionStart	Injects the always-on project summary — top decisions with their rejected alternatives — into every new session.
UserPromptSubmit	Two-tier pre-task veto: tier 1 — POST /veto-scan (deterministic exact match on rejected[], no embeddings); tier 2 — semantic GET /decisions?query= for longer prompts. Warns before Claude starts working.
Stop	Ships the completed turn to Perception for server-side decision extraction (async — never blocks your session). Capture becomes deterministic instead of depending on the model remembering to call MCP tools.

Every hook fails open: if Perception is down or unconfigured, your session is unaffected.

Install

Requires a running RoBrain stack and a registered project:

npx robrain@latest up                  # start Postgres + Perception (no clone needed)
npx robrain init-project               # register this project, warm-start memory

Then add the plugin:

claude plugin marketplace add adelinamart/robrain
claude plugin install robrain@robrain

Connection settings come from ~/.robrain/config.json (written by robrain up / robrain install); PERCEPTION_URL / PERCEPTION_API_KEY environment variables override it.

Notes

• Works alongside the Sensing MCP. Perception deduplicates near-identical decisions server-side, so running both double-captures nothing. With the plugin installed, the mandatory sensing_record_turn block in CLAUDE.md becomes optional — hooks capture deterministically.
• Cross-tool by design. Decisions captured here surface in Cursor, Copilot, and Codex through the same Perception store (robrain install --self-hosted).
• Your data stays yours. Hooks talk only to the Perception URL you configured — by default a Docker container on localhost.