jj-valentine/cerebellum

Capture thoughts from anywhere. Retrieve them semantically. Every AI tool you use — Claude, Cursor, VS Code — connects to the same brain via a single MCP server.

How it works

capture "thought text"
    ↓
embed (text-embedding-3-small)  +  classify (gpt-4o-mini)
    ↓
Postgres + pgvector (Supabase)
    ↓
semantic_search / list_recent / stats  ←  any MCP client

Cost: ~$0.10–0.30/month at 20 thoughts/day.

Stack

Layer	Tech
Storage	Supabase (Postgres + pgvector, HNSW index)
Embeddings	`openai/text-embedding-3-small` via OpenRouter
Classifier	`openai/gpt-4o-mini` via OpenRouter
Protocol	MCP (`@modelcontextprotocol/sdk`)
CLI	Node.js + TypeScript

MCP Tools

Tool	Description
`semantic_search`	Find thoughts by meaning, not keywords
`list_recent`	Retrieve thoughts by time window
`stats`	Totals, type breakdown, top topics & people
`capture`	Write to the brain from any MCP client

CLI

npm run cli -- "I noticed that X tends to happen when Y"   # capture
npm run cli -- search "project planning"                    # semantic search
npm run cli -- recent --days 7 --limit 20                  # recent thoughts
npm run cli -- stats                                        # overview

Setup

1. Clone and install

git clone https://github.com/jj-valentine/cerebellum
cd cerebellum
npm install

2. Configure environment

cp .env.example .env
# Fill in SUPABASE_URL, SUPABASE_SERVICE_KEY, OPENROUTER_API_KEY

3. Run schema in Supabase SQL editor

# Copy contents of schema/schema.sql → paste into Supabase SQL Editor → Run

4. Build and register MCP server

npm run build
claude mcp add --transport stdio --scope user cerebellum -- node /path/to/dist/mcp/server.js

Metadata auto-extracted per thought

Each captured thought is automatically classified into:

Type — observation · task · idea · reference · people · preference
Topics — 1–3 tags
Mentions — mentioned names
Action items — implied next steps

License

MIT