GBrain
GBrain is an open-source personal knowledge management system created by Y Combinator President Garry Tan to provide AI agents with persistent, long-term memory. Designed around a Markdown-first architecture, the system stores knowledge as human-readable files organized under a "compiled truth" and "timeline" structure, allowing agents to read and write information continuously across conversations[^c1]. The system operates on a compounding memory mechanism: each interaction triggers a knowledge base lookup before response generation and writes new knowledge back afterward, building continuous contextual accumulation over time[^c9].
The project was inspired by Vannevar Bush's 1945 Memex concept and Andrej Karpathy's LLM Wiki paradigm, which proposed that AI knowledge should be compiled once and maintained rather than reconstructed on every query[^c2][^c6]. GBrain fills what developers describe as the "missing persistence layer" for AI agents, addressing the limitation that most agents cannot retain knowledge or learn across sessions[^c4]. The project follows a "thin harness, fat skills" architecture that keeps the core runtime minimal while concentrating intelligence in Markdown skill files[^c13].
By May 2026, Garry Tan's personal GBrain installation had grown to over 17,000 documents, 4,300 contacts, and 700 companies[^c3]; the project had accumulated over 16,000 GitHub stars by mid-May, reaching 18,300 stars and 2,500 forks by late May[^c7][^c16]. Notable stargazers included Shopify cofounder Tobi Lutke, Pandas author Wes McKinney, and PingCAP cofounder Ed Huang[^c14]. Tan described GBrain as "the production system I use every day"[^c17]. Community extension projects began appearing within weeks of the initial release, bridging the system with tools such as Notion and using GBrain as a downstream project knowledge repository[^c15]. By late May 2026, a broader ecosystem of community-built second brain systems had emerged, extending GBrain's patterns to multi-agent runtimes, deterministic observing-memory layers, and portable cross-tool knowledge bases.
Later versions introduced code graph support extending the knowledge graph paradigm to software symbols, self-healing diagnostics via automated remediation, and an autopilot mode for continuous background operation[^c10][^c11][^c12]. The system has generated both enthusiasm for its approach to compounding AI memory and debate over its architectural decisions, particularly regarding features that rely on LLM-interpreted instructions rather than deterministic code[^c5].