A personal, database-backed memory system that speaks MCP. Any AI tool — Claude, Cursor, ChatGPT, whatever ships next year — queries the same memory store without integration work. One protocol, every engine.

Raw thoughts don't go straight to the database. Three layers stand between capture and storage.

Quickstart

git clone https://github.com/jj-valentine/cerebellum
cd cerebellum && npm install
cp .env.example .env        # fill in SUPABASE_URL, SUPABASE_SERVICE_KEY, OPENROUTER_API_KEY

Paste schema/schema.sql into the Supabase SQL Editor and run it. Then:

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

# Add to your shell config:
alias memo="node --import tsx/esm /path/to/cerebellum/src/cli/index.ts"

memo "first thought"    # you're in

The prompts/ directory has five ready-to-use prompts for seeding the brain: memory migration, second brain migration, a personalized capture discovery interview, quick capture templates, and a weekly review ritual.

Architecture

memo "thought"
    ↓
Operator  (buffer · synthesis · TTL)
    ↓
Gatekeeper  (quality · contradiction · adversarial review)
    ↓
memo review  (you have final say)
    ↓
embed + classify → Postgres + pgvector (Supabase)
    ↓
semantic_search / list_recent / stats / capture  ←  any MCP client

Operator (Weaver, Mentat, Curator... still deciding — suggestions welcome)

Every capture lands in a short-term buffer before it touches the database. The Operator is an LLM crawling that buffer — picking apart fragments, finding the threads that belong together, synthesizing what can be synthesized. Think less pipeline step, more something alive in the web.

Three calls:

• pass-through — complete, self-contained thought. Route it.
• hold — low-signal fragment. Sit. Wait for the rest.
• synthesise — two or more buffered entries share a theme. Collapse them into one stronger thought. Discard the fragments.

Three half-baked notes about a decision you're wrestling with become one coherent insight by the time they reach the next layer. The fragments never reach the database. The buffer runs on a serialized async chain — concurrent captures don't corrupt each other, and TTL expiry never silently discards. If synthesis fails, entries route individually. Nothing gets lost.

Use memo --axiom "..." to skip the Operator entirely and send something straight to the queue as an axiom.

Gatekeeper

What survives the Operator hits a second LLM evaluation.

The Gatekeeper scores each thought 1–10 (Noise → Insight-grade), runs an adversarial note on borderline items (scores 4–7), checks for contradictions against everything already in the database, and flags veto violations — captures that would contradict a directive you've already marked inviolable.

Output: a recommendation (keep · drop · improve · axiom) and a reformulation if it thinks the thought can be sharper.

Axioms

An axiom is a permanent directive — carved in stone, not written on a whiteboard. Once approved, it doesn't just sit in the database differently; the Gatekeeper actively flags any future capture that would contradict it.

memo --axiom "never ship without a review queue" — skips the Operator, hits your queue marked as axiom, and once you approve it, it's enforced from that point forward. A first-class thought.

You (the Architect)

Nothing reaches the database without your sign-off. memo review walks you through the queue one entry at a time: score, analysis, the skeptic's note if it's borderline, suggested reformulation. Keep it, drop it, edit it, or promote it to axiom. You have final say on everything.

Stack

| Layer | Tech | |---|---| | Storage | Supabase (Postgres + pgvector, HNSW index) | | Embeddings | openai/text-embedding-3-small via OpenRouter | | Classifier | openai/gpt-4o-mini via OpenRouter (configurable) | | Gatekeeper | anthropic/claude-sonnet-4-6 via OpenRouter (configurable) | | Operator | anthropic/claude-sonnet-4-6 via OpenRouter (configurable) | | Protocol | MCP (@modelcontextprotocol/sdk) | | HTTP daemon | Express on 127.0.0.1:4891 | | CLI | Node.js + TypeScript (memo alias) |

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 |

Also available over HTTP at POST /mcp when the daemon is running.

CLI

memo "thought"                        # capture → Operator → GK queue
memo --axiom "directive"              # bypass Operator, queue as axiom
memo review                           # interactive GK queue review
memo web                              # inspect/force-synthesise/discard held entries
memo search "what was I thinking"     # semantic search
memo recent [--days 7] [--limit 20]   # recent thoughts
memo stats                            # overview
memo seed <file.json>                 # batch import from JSON
memo seed --dry-run <file.json>       # preview without writing
memo seed --undo                      # delete all seeded thoughts

Full Setup

Environment

cp .env.example .env
# Required: SUPABASE_URL, SUPABASE_SERVICE_KEY, OPENROUTER_API_KEY
# HTTP daemon: generate CEREBELLUM_API_KEY with: openssl rand -hex 32

Schema — paste schema/schema.sql into the Supabase SQL Editor and run it.

MCP server

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

CLI alias — add to your shell config:

alias memo="node --import tsx/esm /path/to/cerebellum/src/cli/index.ts"

HTTP daemon (optional) — node dist/http/main.js starts on 127.0.0.1:4891. The /mcp endpoint is unauthenticated; /api/* requires a Bearer token. This is what makes external capture surfaces (browser extensions, editor plugins, voice input) possible.

Metadata auto-extracted per thought

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

Cost

~$0.10–0.30/month at 20 thoughts/day. Operator and Gatekeeper both default to claude-sonnet-4-6 but work well with openai/gpt-4o-mini if you want to keep it near zero.

License

MIT