YC CEO releases an open-source production-grade AI memory system “GBrain,” with an original dream cycle mechanism that gives OpenClaw a second brain

Y Combinator CEO Garry Tan announced that he has open-sourced “GBrain,” a production-grade AI Agent memory system he uses for everyday personal use. This “second brain” built specifically for agents like OpenClaw uses an original “Dream Cycle” mechanism and hybrid search with Postgres, enabling AI to achieve perfect global recall (Total Recall).
(Background: Putin ordered the creation of a “Russian autonomous AI model”: the future of sovereignty and survival will depend on artificial intelligence—set the target of achieving nationwide adoption by 2030)
(Background add-on: OpenAI CEO Sam Altman’s residence was hit with a Molotov cocktail! A late-night post reflecting: AGI is like “the One Ring,” and AI power must be democratized)

Table of Contents

Toggle

• Building AI’s “second brain”: a minimalist architecture rooted in pragmatism
• Key highlights: the original “Dream Cycle” and the compiled truth design
• Left hand gstack, right hand GBrain: stitching together mini-AGI

At a time when the AI Agent boom is sweeping across the globe, “Memory System” is being regarded by the industry as a more critical technical bottleneck than the base model itself. To address the pain point that AI is always “turning its head and forgetting,” Y Combinator (YC) President Garry Tan recently generously open-sourced the production-grade AI memory system he is personally using on GitHub — GBrain.

On the X platform, Garry Tan emphasized that GBrain is absolutely not an experimental toy, but a personal knowledge management and memory system (second brain) he is truly deploying in production environments. Its ultimate goal is to help developers build their own “mini-AGI.”

If you want your OpenClaw or Hermes Agent to be able to have perfect total recall of all 10,000+ markdown files, GBrain is here to help.

It’s exactly my OpenClaw/Hermes Agent setup. MIT-licensed open source. Hope it helps you build your mini-AGI.https://t.co/yFpFU4pn5b

— Garry Tan (@garrytan) April 10, 2026

Building AI’s “second brain”: a minimalist architecture rooted in pragmatism

GBrain is mainly tailored for local or controllable AI agents such as OpenClaw or Hermes Agent. The system does away with overly complex SaaS architectures and chooses a “minimal but powerful” engineering implementation path: built on Markdown files and Git Repos at the base layer, plus Postgres as the retrieval layer.

Even though the architecture is simple, the amount of data it carries is quite staggering. As Garry Tan revealed, his GBrain currently indexes more than 10,000 Markdown files and over 5,800 Apple Notes, covering all meeting notes and conversation history. Through Postgres hybrid search (vector retrieval + keyword retrieval) and semantic graph technology, the agent can not only find specific information, but also deeply understand relationships and context between real entities.

Key highlights: the original “Dream Cycle” and the compiled truth design

GBrain is considered a production-grade brain thanks to several breakthrough core innovation mechanisms:

• Dream Cycle: This is GBrain’s most eye-catching feature. Just as humans reorganize memories while sleeping, GBrain will have the AI Agent automatically run every night (or on a regular schedule), scanning all new conversations and content from that day. It will automatically fill in missing entity data, repair broken citation links, and integrate new memories—so the knowledge base can achieve “self-improvement and enrichment.”
• Compiled Truth + Timeline: To ensure the reliability and auditability of AI outputs, GBrain uses a unique data structure design. The “top” of documents stores the latest answers and knowledge (which can be updated and overwritten as new information arrives); the “bottom” preserves a strict chain of evidence (only append, never delete).
• Supports 30+ MCP tools: The system deeply integrates calls to more than 30 Model Context Protocol (MCP) tools, enabling the agent to carry out highly complex query and operation skill packages (Skill pack).

Left hand gstack, right hand GBrain: stitching together mini-AGI

Many developers may confuse GBrain with another project Garry Tan open-sourced earlier, “gstack.” In fact, the two are perfectly complementary:

gstack focuses on “execution capability.” It is the Claude Code virtual engineering team workflow (including skill sets for roles like CEO, engineering manager, QA, and more), helping Garry Tan achieve high-intensity productivity—“writing 600k lines of code in 60 days.” Meanwhile, GBrain focuses on “long-term memory and knowledge management,” acting as the super-brain for the entire team.

Currently, GBrain has been released as fully open source under the MIT license on GitHub (project link). For developers who are exploring real-world AI Agent applications, this highly transparent, hands-on solution personally released by this YC figure is undoubtedly a valuable treasure trove of practical reference value.