GBrain

Overview

GBrain is an open-source retrieval, memory, and knowledge-wiki system created and shared by Garry Tan. Across newsletter mentions, it appears in two closely related forms: as a retrieval layer for AI agents and as a practical GitHub repository that packages reusable “skills” alongside the beginnings of a "Karpathy-style knowledge wiki." It is positioned as MIT-licensed infrastructure for building personal or organizational memory systems that can plug into agent stacks such as OpenClaw, Hermes Agent, Codex, and Claude Code.

For AI Product Managers, GBrain matters because it sits at the intersection of three strategic themes: agent memory, reusable workflows, and knowledge management. Rather than treating retrieval as a generic RAG add-on, GBrain is framed as a production-oriented system for giving agents access to large bodies of markdown knowledge, structured skills, and company context. That makes it relevant for PMs evaluating how to operationalize internal knowledge, improve agent task performance, and design scalable “company brain” workflows that support product, engineering, research, and client-facing teams.

Key Developments

• 2026-05-18: Garry Tan announced that GBrain adopted ZeroEntropy as its recommended default embedding and re-ranking engine, replacing OpenAI and Voyage AI.
• 2026-05-20: GBrain was introduced as an MIT-licensed open-source retrieval and memory system that achieved state-of-the-art results on LongMemEval, reportedly outperforming known open-source repositories by more than 1% without LLM query rewriting.
• 2026-05-24: Garry Tan launched GBrain as a state-of-the-art retrieval engine for agents, built for OpenClaw and Hermes with MCP server support for broader agent-harness compatibility. Around the same time, GBrain updates added synthesized answers in addition to retrieval, with internal comparisons between “GBrain Search” and “GBrain Think.”
• 2026-06-01: GBrain was open-sourced on GitHub under the MIT license, with Garry Tan outlining a roughly 30-minute setup using a 350k-page markdown LLM wiki together with an OpenClaw/Hermes agent.
• 2026-06-06: GBrain was described as a modular “company brain” framework that structures work through scoped AI agents organized into client pods, signaling a broader architecture beyond retrieval alone.
• 2026-06-16: Garry Tan introduced GBrain as an open-source retrieval layer for AI agents, comparing it to a just-in-time “3-book librarian” and noting adoption by thousands of developers.
• 2026-07-21: GBrain was formally launched as a free open-source retrieval library optimized for Hermes Agent and OpenClaw, with Codex and Claude Code support, and positioned as infrastructure powering Garry Tan’s personal company brain and AI.
• 2026-07-28: Garry Tan shared that GBrain had reached state-of-the-art performance for agentic retrieval without LLM rewriting, with evaluation details published in the gbrain-evals GitHub repository.
• 2026-08-13: GBrain v0.45.6.0 was released with 17 new brain skills, reportedly hardened through use with Garry Tan’s personal OpenClaw agent across hundreds of thousands of markdown files. Support for Codex and Claude Code was emphasized.
• 2026-08-18: Garry Tan highlighted the GBrain GitHub repository as a private-repo-friendly package containing 70 proven skills and the early version of a “Karpathy-style knowledge wiki,” describing it as free, open source, and fully documented in the README.

Relevance to AI PMs

1. Designing practical memory systems for agents: GBrain offers a concrete example of how to move beyond simple chatbot retrieval toward persistent agent memory built on markdown knowledge bases, reusable skills, and evaluation-backed retrieval. PMs can use it as a reference architecture when defining internal copilots, research assistants, or product-ops agents.

2. Building a reusable knowledge layer for teams: The emphasis on a knowledge wiki plus skill repository is especially relevant for PMs trying to standardize best practices across product, engineering, support, and go-to-market workflows. It suggests a way to encode playbooks, decision logs, onboarding docs, and task recipes into an agent-accessible system.

3. Evaluating retrieval infrastructure choices: GBrain is repeatedly framed around measurable retrieval quality, embedding/reranking choices, and support for multiple agent environments. AI PMs can study it to better assess tradeoffs among RAG stacks, MCP-compatible tools, embedding vendors, and internal “company brain” strategies before committing roadmap or platform resources.

Related

• garry-tan: Creator and primary public advocate of GBrain, sharing product updates, evaluation claims, and architecture examples.
• github: Main distribution channel for the open-source repository, documentation, and associated evaluation artifacts.
• openclaw and hermes-agent / hermes: Core agent environments GBrain is repeatedly described as being optimized for.
• claude-code and codex: Developer-agent tools that GBrain later added or emphasized support for, expanding its practical compatibility.
• gbrain-evals: The companion evaluation repository used to support claims about retrieval performance and state-of-the-art results.
• rag: GBrain is relevant as an evolution of retrieval-augmented generation toward stronger agent memory and retrieval quality, including cases without LLM query rewriting.
• zeroentropy, openai, voyage-ai: Referenced through GBrain’s embedding and reranking stack choices, with ZeroEntropy later becoming the recommended default over OpenAI and Voyage AI.
• mcp: Important because GBrain was presented as having MCP server support, making it easier to integrate into diverse agent harnesses.
• company-brain and client-pods: Concepts tied to the broader organizational architecture around GBrain, where scoped agents and structured knowledge power workflows.
• karpathy-style-knowledge-wiki: A useful shorthand for the wiki-like, markdown-heavy knowledge organization approach associated with the repository.
• git-history and dry: Related conceptually to reusable skills, versioned knowledge, and maintainable operational playbooks inside a repository-based brain.
• gstack, opus, hyojun_at: Adjacent entities in the broader ecosystem and conversation around agent tooling, knowledge systems, and AI workflow design.