Open Brain 🧠

Built on Nate B. Jones's Open Brain architecture. This is MonkeyRun's implementation and extension of his open-source personal memory system — we've added Gmail capture, RAG chunking for long-form content, and documented the journey of building it with AI.

MonkeyRun is a software studio that builds and ships products using AI as the primary development partner — using tools like Cursor, Claude, and OpenClaw to manage real projects end-to-end. This repo is one of those experiments: building a persistent AI memory layer and learning in public.

Open Brain is an MCP server that lets AI agents save, search, and retrieve "thoughts" using semantic vector search. You capture ideas, decisions, emails, and notes — and any AI client you use (Claude, ChatGPT, OpenClaw) can search them by meaning, not just keyword.

Data Sources

Source	How	Status
Discord #capture channel	Push — type into channel	✅ Built-in
Gmail (sent + starred)	Pull — run scripts/pull-gmail.ts	✅ Available
ChatGPT export	Bulk import — run scripts/import-chatgpt.py	✅ Available
Manual (any AI client)	capture_thought MCP tool	✅ Built-in

ChatGPT Export Importer

Turn your ChatGPT data export into searchable memories. Processes your entire conversation history, filters out low-signal content (quick lookups, one-off creative tasks), and distills lasting knowledge into Open Brain.

# Export your data: chatgpt.com → Settings → Data Controls → Export Data
# Then run:
pip install requests
python scripts/import-chatgpt.py ~/Downloads/chatgpt-export.zip --dry-run --limit 10
python scripts/import-chatgpt.py ~/Downloads/chatgpt-export.zip

What it captures: decisions and reasoning, people with context, project plans, lessons learned, business context, personal values
What it skips: poems, jokes, one-off writing tasks, generic Q&A, coding help with no lasting context

Each ingested thought stores:

• A distilled summary (used for semantic search)
• The original conversation text (retrievable on demand)
• A direct link back to the source (https://chatgpt.com/c/<id>)

Supports OpenRouter (default, ~$0.15/100 conversations) or local Ollama for zero-cost inference. See scripts/import-chatgpt.py --help for all options.

Guides & Extensions

Guide	Description
Email Capture — Visual Overview	Add Gmail to your persistent memory — pull-based ingestion, RAG chunking, the five things that broke
Email Capture — Full Guide	Step-by-step setup, troubleshooting, automation, and security notes
Substack Post Draft	Community contribution post for Nate B. Jones's Open Brain series

Architecture

• Database: Supabase PostgreSQL + pgvector (1536-dimension embeddings, full_text column for source storage)
• Embeddings: OpenRouter → openai/text-embedding-3-small
• Metadata extraction: OpenRouter → openai/gpt-4o-mini
• MCP Server: Deno Edge Function (Supabase) using @modelcontextprotocol/sdk
• Capture: Discord #capture channel → ingest-thought Edge Function

MCP Tools

Tool	Description
capture_thought	Save a thought with auto-embedding + LLM-extracted metadata
search_thoughts	Semantic search — returns source URL and optional full original text
list_thoughts	Browse recent thoughts with filters (type, topic, person, days)
thought_stats	Totals, type distribution, top topics, top people

Deployment

supabase functions deploy open-brain-mcp --no-verify-jwt
supabase functions deploy ingest-thought --no-verify-jwt

Infrastructure

• Supabase Project: Your own Supabase project (see supabase/config.toml after supabase link)
• Auth: Custom key-based (x-brain-key header or ?key= query param)
• JWT: Disabled at function level (self-managed auth)