YC CEO Garry Tan open-sources an AI memory system called GBrain: making AI assistants smarter with every conversation

Y Combinator’s current CEO, Garry Tan, has open-sourced an AI agent knowledge system called GBrain on GitHub, enabling AI assistants to build up lasting long-term memory. This is not a proof of concept—it’s a productivity tool Tan personally uses in his day-to-day work. He has accumulated over 10,000 Markdown files through an OpenClaw agent, covering every person, company, and concept he’s come into contact with.

Why does the YC CEO build a knowledge management tool himself?

Tan’s core insight is this: today’s AI agents start from zero every time you have a conversation—they don’t remember who you are, don’t know what you’re doing, and don’t understand your context. The problem GBrain aims to solve is to make the agent smarter with every conversation, rather than having it act like it’s the first time you’ve met.

The operating logic is a continuous cycle of “read → respond → write”: when the agent receives a message, it first detects the entities involved (people’s names, companies, concepts), then looks up the relevant knowledge already in GBrain, responds with complete context, and finally writes the new information learned from the conversation back into the knowledge base. Each interaction accumulates, and the results grow with compounding effect over time.

Knowledge structure: compile truth plus a timeline

GBrain’s knowledge storage format is quite unique. Each entity (person, company, concept) has its own dedicated page, made up of two parts:

“Compiled Truth” is your best current understanding of that entity, and it gets rewritten as new evidence appears. “Timeline” is purely an append-only record of evidence—it only grows and never changes, documenting each interaction, every information source, and each timestamp.

This design makes knowledge traceable: you don’t just know what something is—you can also trace when you learned it and where it came from.

Data sources: meetings, Email, Twitter, phone calls—fully automated imports

GBrain offers multiple ways to automate integrations so knowledge automatically flows into the system:

Integration Source Function
Gmail Automatically converts email content into an entity page
Google Calendar Daily schedules into searchable knowledge pages
Twitter / X Timeline, mentions, and deletion tracking
Voice calls Transcribed into knowledge pages via Twilio + OpenAI Realtime
Meeting notes Circleback Automatically converts transcripts into brain pages

Technical architecture: build a complete knowledge base in 30 minutes

GBrain defaults to using PGLite—a lightweight Postgres 17.5 that runs through WebAssembly. No database server setup is required, and it can be started in under two seconds. Search uses a hybrid mode, combining vector semantic search (OpenAI embeddings) and keyword search, with the two result sets integrated via Reciprocal Rank Fusion.

The system supports three ways of using it: a command-line tool (CLI), an MCP server (directly connectable to tools like Claude Code, Cursor, etc.), and a TypeScript library for developers to integrate. In MCP server mode, it provides 30 tools, including page read/write, search, knowledge graph traversal, and file uploads.

The significance for the AI agent ecosystem

The introduction of GBrain addresses a core problem in the AI agent field: memory. Although most mainstream AI tools (Claude, ChatGPT) include basic memory features, they are mostly limited to conversation preference levels. GBrain proposes a more ambitious vision—giving agents structured “world knowledge,” not just remembering that you prefer to use Traditional Chinese.

In the documentation, Garry Tan specifically distinguishes three layers of memory: “world knowledge” managed by GBrain (people, companies, meetings, concepts), the agent’s own “operational state” (preferences, decision-making, behavioral patterns), and real-time “conversational context.” He believes that when an AI agent runs, it should check all three layers together to deliver truly personalized service.

This system comes from one of Silicon Valley’s most influential venture capitalists. He uses it every day to manage interactions with hundreds of founders and investors. When the YC CEO believes AI agents need this kind of knowledge infrastructure, that in itself is a signal worth paying attention to.

This article, “YC CEO Garry Tan Open-Sources AI Memory System GBrain: Making AI Assistants Smarter Every Conversation,” was first published on Chain News ABMedia.