YC CEO Open-Sources His Second Brain: The Most Important Thing Across the 8 Repos Isn't the List

Ryven listed 8 second-brain repos on X, and Eric Hartford retweeted it. The hook in the retweet is blunt: YC President and CEO Garry Tan has open-sourced his second brain, and the repo is called gbrain.

The original post included a demo video:

These tools solve a specific problem: an agent loses its memory every time it starts a new session. You drop articles, meeting notes, and code decisions into a chat window, and it can answer at the time, but once you close the tab, they're gone. RAG goes back to the original files on every question, which suits one-off retrieval but not long-term accumulation. The LLM Wiki Karpathy proposed in a gist takes a different path: have the model compile the material into a set of interlinked markdown pages. When new material comes in, it updates the old pages; questions are answered from the wiki, and good answers are written back into the wiki.

Karpathy splits the structure into three layers. raw holds original material, which the model only reads and never changes. wiki holds model-generated summaries, entity pages, concept pages, and cross-links; humans read it, and the model writes it. schema is a rules file that tells the model how to organize directories and how to do ingest and query. The operations are simple too: ingest compiles new material into the wiki, query synthesizes answers from the wiki, and lint regularly checks for contradictions, orphan pages, missing links, and outdated conclusions. index.md serves as the content index, and log.md records what happened over time. Obsidian is the browser here, the model is the maintainer, and the wiki is a place that can be lived in long term.

Garry Tan's gbrain turns this approach into an engineering system. The README says it runs his OpenClaw and Hermes deployments, at a scale of 155,795 pages, 24,589 people, 5,340 companies, and 66 cron jobs. It stores explicit facts and preserves provenance, supports corrections and retractions, and the same memory can be used by different agents. Retrieval starts with keyword search; semantic search, synthesized answers, and knowledge graphs are add-on options. gbrain think returns answers with citations and also flags which memories are stale or lack sources. The install commands are not complicated:

bun install -g github:garrytan/gbrain#stable
gbrain setup claude-code
gbrain init --pglite --no-embedding

The README specifically warns not to install gbrain with npm; the package with the same name on npm is not the same thing. The project is MIT-licensed, can be self-hosted, and also has a paid hosting option.

ai-memory takes another path. It makes cross-agent handoff the core: Claude Code exits halfway through, Codex picks up in the same directory, and there is no need to re-explain the architecture, failed approaches, and open questions. It also uses a git-backed markdown wiki as the source, SQLite as a derived index, default zero LLM calls, and supports more than 20 harnesses. A handoff has an owner and is claimed only once. This design is more like shared working memory for multiple coding agents than a single chat knowledge base.

The 8 repos in the original post can be viewed by entry point:

• gbrain: Garry Tan's personal second brain; facts carry provenance and are shared across agents.
• llm wiki: Karpathy's method file; the three-layer structure and ingest, query, and lint are all here.
• ai-memory: long-term memory across coding agents; markdown is the source, and handoffs have a protocol.
• agent second brain: send voice notes in Telegram and they become linked notes in Obsidian.
• cog second brain: 33 skills and 10 agents run on pure markdown, no database.
• open second brain: nightly jobs distill rules you repeatedly correct into agent rules.
• claude second brain: one npx command sets up Obsidian, Claude Code, and local search.
• baalda: team second brain; humans and agents edit the same markdown files.

With the list laid out, it is easy to fall into an illusion: install one and you have a second brain. The real barrier is maintenance. The raw directory must be immutable, the wiki must be linted regularly, rule files must not balloon, and permissions must be tightened. Anthropic's docs say CLAUDE.md is context, not enforced configuration. To block deletions and force pushes, you need permissions or a PreToolUse hook, not just a line asking the model not to delete. You also need backups; git or a sync drive lets you undo bad edits.

Privacy and cost are just as real. When Claude reads files, it sends their contents to the model, so don't put sensitive material in raw; Pro and Max quotas are shared by Claude and Claude Code, and importing hundreds of files the first time can burn through them quickly—start with 5 to 10 sources. In the comments on Karpathy's gist, some also question it: the ingest cost of compiling a wiki may exceed the savings, and when questions are unpredictable, the wiki easily lacks detail, so agents still have to go back to the original sources. That caveat is worth keeping. A second brain suits domains with a clear direction for accumulation; it is not for dumping your bookmarks in all at once and waiting for it to become smart.

You don't need to finish reading all 8 repos to get started minimally. Take one source you will genuinely revisit, create raw, wiki, and output, write a short set of rules, run one ingest, ask a cross-source question, and save the good answer back into the wiki. The next day, add five more sources. Do a lint once a week. Get this loop working first, then decide whether to use gbrain, ai-memory, or build your own. Collecting eight repos does not produce a second brain. Getting one source working does.