The CEO of Y Combinator built a personal AI knowledge system — a knowledge graph, 74 MCP tools, and a nightly consolidation loop. Then he published it so anyone can run it.
Out-of-the-box LLMs are stateless. They re-meet you every session, lose context across days, and can't track the entities, relationships, and history that make your judgment yours. For someone like Garry Tan — running an accelerator, talking to hundreds of founders — that's a non-starter.
GBrain is three layers stacked on top of an off-the-shelf model. The model didn't get smarter — the context around it did.
GBrain runs 24/7. The knowledge graph compounds. Citations get fixed. Entities get merged. By morning, the AI knows more — and more accurately — than it did yesterday. And because it's open source, you can run the same thing.
Garry also published gstack — his 23-tool Claude Code configuration. Same posture: keep the model fixed, sharpen the context and tools around it.
Garry didn't fine-tune anything. He didn't train his own model. He gave the model persistent memory, real tools, and a nightly consolidation loop. That's the gap between AI that forgets you and AI that gets sharper every day.