GBrain: Garry Tan's Open-Source AI Agent Brain With Self-Wiring Knowledge Graph

AIDeveloper44 Team

What Is GBrain?

GBrain is an open-source AI agent memory system built by Garry Tan — President and CEO of Y Combinator — to run his actual AI agents. It's not a demo or prototype: the production brain currently manages 146,646 pages, 24,585 people, 5,339 companies, and 66 cron jobs running autonomously.

The core thesis: your AI agent is smart but forgetful. GBrain gives it a brain that wires itself.

How the Knowledge Graph Self-Wires

Every page write to GBrain automatically extracts entity references and creates typed links — attended, works_at, invested_in, founded, advises — with zero LLM calls. The graph builds itself as the agent ingests content, meaning:

• Ask "who works at Acme AI?" and get answers vector search alone can't reach
• Ask "what did Bob invest in this quarter?" and get structured timeline results
• Backlink-boosted ranking surfaces the most-connected knowledge automatically

Benchmark Performance

GBrain has been benchmarked against standard retrieval baselines on a 240-page Opus-generated rich-prose corpus:

• P@5: 49.1%, R@5: 97.9% — full hybrid graph mode
• +31.4 points P@5 over graph-disabled variant
• Outperforms ripgrep-BM25 + vector-only RAG by a similar margin

The default embedding model as of v0.36.2.0 is ZeroEntropy (zembed-1 at 1280d via Matryoshka), which is 2.2× faster and 2.6× cheaper than OpenAI embeddings while winning 11 of 20 head-to-head queries.

Autonomous Brain Maintenance

GBrain v0.36.4.0 introduced an autonomous health optimization loop:

gbrain doctor --remediate --yes --target-score 90 --max-usd 5

This single command computes a dependency-ordered remediation plan, submits each phase as a background job, re-checks the brain health score between steps, and refuses to exceed a user-defined cost cap. Cron can run it unattended nightly.

Installation

GBrain installs in under 30 minutes and runs in three modes: as a skillpack scaffold for agent platforms (OpenClaw, Hermes), as a standalone CLI, or as an MCP server for any MCP-compatible client (Claude Code, Cursor, Windsurf). PGLite mode requires no server or Docker — database ready in 2 seconds.

bun install -g github:garrytan/gbrain
gbrain init --pglite
gbrain doctor

Full documentation at github.com/garrytan/gbrain.

Enjoyed this?

Get more posts like this delivered to your inbox.

Subscribe