Brain

Brain is a persistent, outcome-grounded memory system I built for Claude Code. It gives an AI agent long-term memory that learns from what actually worked: every stored memory carries a Bayesian confidence that rises when a decision it informed proved correct and falls when it didn't. Brain runs as a local FastAPI service backed by per-project SQLite databases, and it exposes itself to Claude Code through an MCP server.

The core mechanism is a background consolidation cycle I called "sleep" — eight phases (replay, contradiction evaluation, pattern extraction, connection formation, homeostasis, promotion) that run on cron and mirror how the mammalian brain reorganizes memory overnight. This was built independently before equivalent features shipped in Claude Code — the commit history is public.

I used Brain to run the pre-registered cosmology study at github.com/EdonZo/ccc-hawking-points, where persistent memory across weeks of analysis sessions was load-bearing rather than convenient.