GBrain is an open-source 📝Personal Knowledge Management system designed for AI agents, created by 📝Garry Tan, president and CEO of 📝Y Combinator. Built in TypeScript on Bun, it transforms meetings, emails, tweets, calendar events, and voice calls into a searchable markdown knowledge base that agents read before every response and write to after every conversation. The system uses a compiled-truth-plus-append-only-timeline structure per page, hybrid search via Postgres with pgvector, and automatic entity detection with backlinking — enabling what Tan describes as a compounding knowledge loop where each agent interaction enriches the base for the next.
GBrain supports multiple deployment modes: PGLite for zero-config local use, Supabase for production scale, and self-hosted Postgres. It exposes over 30 MCP tools for integration with 📝Claude Code, 📝Cursor, and Windsurf, alongside a CLI with 37 operations covering import, search, query, sync, and embedding. Data ingestion pipelines include 📝Gmail, 📝Google Calendar, 📝Twilio voice calls (with 📝OpenAI Realtime), 📝X (Twitter), and Circleback meeting transcripts. A "dream cycle" feature runs nightly cron jobs for entity sweeps, citation fixes, and memory consolidation. The project requires frontier-class models — tested with Claude Opus 4.6 and GPT-5.4 Thinking — and has accumulated 7,500+ stars on GitHub since release.
The architecture maps closely to the pattern 📝Andrej Karpathy described in his 📝LLM Knowledge Bases post: raw sources compiled into interlinked markdown, an agent-readable schema file, and outputs filed back into the wiki to compound over time. GBrain is a concrete implementation of that thesis, packaged as a developer tool rather than an end-user product.