Wink Pings

YC CEO Just Open-Sourced His Second Brain, and Most People Don't Know Yet

YC CEO Garry Tan has open-sourced gbrain, the second brain he uses himself: every fact carries its source, and all agents share a single memory. This piece rounds up 8 related repos and the LLM Wiki thinking behind them.

Y Combinator (YC) CEO Garry Tan has put his second brain on GitHub. MIT license, free to self-host. After the tweet went out, most people still don't know about it.

A second brain isn't mysticism. It's just a folder of plain-text notes that AI reads, organizes, and cross-links for you. You just keep dumping things in, then ask it directly.

Ryven rounded up 8 repos on X, all circling the same question: AI agents lose their memory every time a session ends, so how do you give them a persistent memory?

(Video from Ryven's original tweet)

This whole line of thinking starts with a gist Andrej Karpathy posted in April 2026: LLM Wiki. It argues against having the model re-read the raw documents every time you ask a question, and instead has the model incrementally maintain a wiki. The source material stays immutable, the wiki is written by the model, and a rules file (such as CLAUDE.md) tells the model how to maintain it. Karpathy's core point: knowledge is compiled once and kept up to date, so you don't re-derive it on every query.

The 8 repos are:

  1. gbrain (GitHub): Garry Tan's own production-grade second brain. Every fact carries its source, and all agents share a single memory. The README says his deployment has already run through 155,795 pages, 24,589 people, and 5,340 companies, with 66 cron jobs running automatically at night. Supports Claude Code, Codex, Cursor, ChatGPT, Grok Bot, and more. gbrain.io is optional paid hosting; the code itself is MIT.

  2. llm wiki (GitHub): where the whole pattern started. Karpathy wrote it up as a gist, and many later implementations grew out of it.

  3. claude-mem (GitHub): runs an observer agent alongside Claude Code that records what happened during a session and automatically brings the important context back in the next session. Supports Claude Code, Codex, Cursor, OpenCode, and more.

  4. agent second brain (GitHub): send a voice message on Telegram and it automatically becomes a bidirectional-linked note in Obsidian.

  5. supermemory (GitHub): a memory layer for agents that remembers you across tools.

  6. basic memory (GitHub): gives Claude memory using plain Markdown files that you can open and edit in Obsidian.

  7. qmd (GitHub): searches all your Markdown notes locally so agents can find the relevant content in seconds.

  8. obsidian skills (GitHub): teaches agents how to read and write Obsidian notes the right way.

If you want to get hands-on, Ryven also wrote a complete 10-step build guide: Build a Second Brain with Claude Code and Obsidian. The first 5 steps take about 20 minutes; the core structure is raw/ for source material, wiki/ for AI-written pages, output/ for finished work, and CLAUDE.md as the rulebook.

The comments are more interesting than the list. Someone said: memory is only useful when it can be traced back to its source and corrected; otherwise it just makes the agent repeat its mistakes with more confidence. Others said they built the same system but never open Obsidian, doing all their retrieval through the command line. Someone else asked whether it could be built with Grok — at least gbrain officially documents how to connect Grok Bot.

Garry Tan says in the README that the point of building a 150,000-page brain is to treat it as a strategic moat. Then he open-sourced the code. The moat isn't the tool; it's your own data.

Original tweet: Ryven

发布时间: 2026-11-22 11:21