Y Combinator president Garry Tan has open-sourced GBrain, the production-grade AI agent memory system he says he uses in daily work. Built for agents such as OpenClaw and Hermes Agent, the project is framed as a “second brain” for AI, with the stated goal of helping developers assemble their own mini-AGI stack.
A minimal stack built on Markdown, Git repos, and Postgres
Instead of using a heavy SaaS architecture, GBrain takes a simpler route. Its foundation is Markdown files and a Git repo, with Postgres used as the retrieval layer. The search system combines vector retrieval and keyword retrieval, while a semantic graph is used to map relationships and context across entities rather than returning isolated records.
According to Garry Tan, GBrain currently indexes more than 10,000 Markdown files and over 5,800 Apple Notes entries, along with all meeting notes and conversation history. The dataset is large, but the design choice is intentionally restrained: fewer moving parts, more direct control over memory and traceability.
Dream Cycle runs periodic memory cleanup and enrichment
The standout mechanism is Dream Cycle. On a nightly or scheduled basis, the agent scans new conversations and content from the day, fills in missing entity data, repairs broken references, and merges new memory into the knowledge base. It is not just storage. It is an automated reorganization layer for memory.
GBrain also uses a Compiled Truth + Timeline structure. The top portion of a document stores the latest answer or knowledge state and can be updated as new information arrives. The lower portion preserves an append-only evidence chain. In the source material, this structure is presented as a way to improve reliability and auditability in AI outputs.
Different from gstack, focused on long-term memory
Tan also drew a distinction between GBrain and his earlier open-source project, gstack. gstack is centered on execution, described as a virtual engineering team workflow for Claude Code with role-based skills such as CEO, engineering manager, and QA. GBrain, by contrast, is built for long-term memory and knowledge management, acting as the memory core for an agent or a broader software team.
The project also integrates 30+ MCP tools, allowing agents to run more complex queries and skill-pack operations. GBrain has been released on GitHub under the MIT license. Based on the source, Tan presents it not as an experiment, but as a system already used in a live production setting for personal knowledge management.