YC's CEO runs the company with 150,000 pages of notes, and this system is now open source

Garry Tan is the CEO of Y Combinator. He put a repository called gbrain on GitHub, and its README lists a string of numbers: 155,795 pages, 24,589 people, 5,340 companies, and 66 cron jobs running while he sleeps. He calls it the "production-grade brain" behind his AI agent.

Ryven surfaced this in a tweet on November 8 and rounded up a list of open-source repos with the same idea. The video shows a bidirectional-link note-taking interface; the idea comes directly from a gist Andrej Karpathy published on April 4, 2026.

That gist is called llm-wiki.md, and its core point is brief: most people still have LLMs read documents via RAG, retrieving and stitching things together from raw files on every question, so knowledge doesn't accumulate. Karpathy suggests a different approach: have the model compile raw material into a continuously maintained wiki; when new material comes in, update entity pages, concept pages, and summaries, and flag contradictions along the way. Later questions query this wiki directly.

Three-layer structure: the raw directory holds raw material, read-only; the wiki directory holds pages written and maintained by the model; plus a schema file (called CLAUDE.md in Claude Code and AGENTS.md in Codex) that defines conventions and workflows. Three operations: ingest feeds in new material, query asks questions, lint runs a health check.

The interface shown in the video:

Ryven listed 8 repositories under different authors:

• gbrain (garrytan): the YC CEO's own brain, with a source kept for every fact and multiple agents sharing the same memory.
• llm wiki (karpathy): the original gist above.
• ai-memory (akitaonrails): lets Claude Code hand off to Codex mid-task without losing context, with memory stored in plain Markdown.
• agent second brain (smixs): send a voice message in Telegram, and it lands in Obsidian as a linked note.
• cog second brain (huytieu): 33 skills, 10 agents, running on plain Markdown with no database.
• open second brain (itechmeat): runs overnight and turns corrections you've made repeatedly into rules agents must follow.
• claude second brain (jessepinkman9900): one npx command sets up Obsidian, Claude Code, and local search.
• baalda (naveedharri): a team second brain where humans and agents edit the same batch of Markdown files.

But the comments were not all praise. In the replies to Karpathy's gist, a user named hyunsoo36 raised a very specific issue: the ingest cost of compiling the wiki is hard to absorb. In his testing, the compiled wiki often lacked detail, and when asking questions the agent still had to go back through the raw material. So this intermediate artifact became an extra navigation layer, adding LLM calls, tool calls, latency, and tokens, while the retrieval cost saved did not make up for it. He suspects this is a mismatch between "synthesis at compile time" and "open-ended future questions," and notes that this is his own usage observation, not a controlled experiment.

Under Ryven's tweet, Sajjad Khan asked another question: in the team version, how do you handle conflicts when two agents edit the same note at the same time? He said he needs visible diffs and source links, so the next agent doesn't treat the merged version directly as fact. Someone else left only "AI slop."

These doubts don't negate the whole idea, but they point out its boundaries. The bottleneck of a second brain has never been collection; it's maintenance and verification. Karpathy himself wrote in the gist that people abandon wikis because the maintenance burden grows faster than the value, while models don't mind the tedium and can edit 15 files at once. The premise is that the model actually gets the edits right. gbrain's countermeasure is a synthesis layer plus citation checks: every assertion carries its source page, and it also lists "the parts the brain doesn't know yet."

If you just want to try it, you don't have to clone eight repos at once. Ryven's original post gives a 30-day plan; the first five steps take about 20 minutes and only require a Claude Pro or Max subscription, Obsidian, and a file you want to save. Steps 6 through 10 make it smarter: add project rooms, save repeated workflows as skills, connect a calendar, and set up scheduled tasks. The original also lists four common failure modes: editing raw directly, turning CLAUDE.md into a monster, asking questions that are too broad, and skipping regular health checks.

Setting it up is only the beginning; the real cost is making it a little more accurate every month than the month before.