Garry Tan releases gbrain, a self-updating memory system for AI agents
YCombinator CEO Garry Tan open-sourced gbrain under MIT license, a production memory architecture that auto-updates markdown knowledge bases, runs hybrid search across vectors and graphs, and builds relationship networks without LLM calls.

"This is what a production brain for agents should look like," observers might say of gbrain, the memory system YCombinator CEO Garry Tan released this week under an MIT license. The tool powers Tan's personal agents OpenClaw and Hermes and treats knowledge as a living database that enriches itself over time.
Gbrain organizes information as markdown files. Each person, company, concept, or meeting gets its own file containing an auto-updated summary and a chronological log of mentions. When new data arrives, the summary rewrites itself; the log appends. Standard agents either forget context between sessions or dump notes into a primitive MEMORY.md file. Gbrain offers multi-layer memory depth by design.
Search runs in parallel across three modes: vector similarity, keyword matching, and graph traversal. The relationship graph — who works where, who attended which meeting, who invested in what — builds itself automatically using regex and rule-based extraction, with no LLM calls required. A single RESOLVER.md file defines 29 skills that keep the system current. One example: meeting-ingestion processes transcripts overnight to enrich participant files.