Garry Tan, President of Y Combinator, released GBrain v0.45.6.0 on August 12. This open-source project adds a persistent memory layer to his AI agent, with 17 new skills added. Tan says these skills were refined on his own OpenClaw agent using over 150,000 pages of markdown files. GBrain now supports Codex and Claude Code, two widely used AI coding assistants.

What problem does GBrain solve? Put simply: AI agents lose all memory once a conversation ends. The decisions you discussed with it last week and the knowledge it learned might be forgotten the next time you open it, or it will confidently make up a wrong answer. GBrain acts as a persistent memory layer that agents can read from and write to, allowing them to accumulate knowledge across sessions. It works out of the box with zero configuration, spinning up a local database with a single command—no cloud services or API keys required.

But Tan emphasizes that GBrain is not a note-taking app or just a vector database. Beyond storing memories, it also has built-in “self-governance” capabilities. The new version is built entirely around the core value of “trust”. Features include: correction-pipeline: if you point out an incorrect fact, it will trace the error to its source and fix it, so the mistake won’t pop up again; fact-check: before delivering a conclusion, it re-verifies information through an independent path, and doesn’t self-grade the answer it writes; data-loss-gate: it shows you the consequences before you delete irreversible data; conversation-archive: it can import years of ChatGPT and Claude conversation history into searchable memory, and automatically strips out API keys; citation-graph-ingest: it builds typed knowledge graphs, enabling search that can reason “what is covered under this topic” instead of just returning a messy pile of vector results; skill-autobench: it generates benchmarks for skills based on real-world usage, getting more accurate the more you use it. It comes with 69 built-in skills in total.


The announcement tweet sparked a lot of discussion. Some netizens commented that the fact that skills were “refined through personal agent usage” is the key point here—most skill libraries are written from imagination, not coming from a system that has to live with its own outputs. Others approve of the “trust-based self-governance” approach, noting that it’s far more dangerous for an agent to remember wrong information than to forget it. But some raised questions: if answer generation and answer verification pull from the same retrieval pool, a bad chunk will contaminate both paths, unless the independent path really queries separate sources. Others went even further: “a folder full of markdown files isn’t personal AGI, it’s just self-indulgent search.” Additionally, some users worry that after crossing 10,000 documents, retrieval noise will overwhelm the signal, and “personal context” will become too generalized.

GBrain positions itself as a “brain layer”, not just another chatbot. It is genuinely delivering solid engineering work: cross-session memory, self-correction, typed knowledge graphs. But the “personal AGI” branding does tend to shift attention away from its engineering advances to conceptual debates. Maybe “personal memory infrastructure” would be a more accurate name. Still, for an open-source project, sparking this level of discussion already proves it has tapped into a real pain point.

Note: GBrain is not published on npm. To install, you need to pull the code directly from GitHub to avoid accidentally installing an unrelated package with the same name.

Project repository: https://github.com/garrytan/gbrain