Garry Tan's Second Brain Has 155,795 Pages and a Correction Button

The number in the README is 155,795 pages. Then 24,589 people, 5,340 companies, and 66 cron jobs running without anyone watching them. It belongs to Garry Tan, president and CEO of Y Combinator, and the software is his, published under MIT with a paid hosted option if you would rather not run a database.
The repository is gbrain, and its one-line pitch is "give the agent you already use a memory you control." That sentence is doing a lot of work, because there are at least three products hiding in it: a memory store, a retrieval system, and a daemon that keeps ingesting while you sleep. The interesting one is not the biggest number.
What "memory" usually means, and what it means here
Most agent memory, in practice, is a vector index with a friendly name. You paste documents in, it embeds them, it retrieves ten chunks that mention your query, and the model stitches something together. It works until you ask a question where two chunks contradict each other, or a fact changed six months ago, or someone needs to know where a claim came from.
GBrain stores explicit facts with their sources. It supports corrections and withdrawal — you can mark something as wrong, and the wrongness propagates instead of sitting in an embedding forever. The same memory is available across different agents rather than being welded to one harness. You start with keyless keyword memory and retrieval, then add semantic search, synthesis and background enrichment when you need them. That ordering is deliberate: the cheap mode works with no model calls at all.
The synthesis layer is the part that changes behavior. Instead of returning ten chunks, it returns an answer with citations and an explicit note about what the brain does not know yet. That gap analysis is unusual. Most retrieval systems are structurally incapable of saying "I have nothing on this," because similarity search always returns its top k, however irrelevant.

Provenance is the actual feature
Tan's framing is that the brain is a moat: never lose context, query what is in your own head without re-reading it. Fair enough, but the moat only holds if the memory is trustworthy, and trust in a memory system comes from two properties that are boring to demo and expensive to build.
The first is attribution. Every fact points at a source. When the agent says a company raised a round, you can walk back to the meeting note or the email that said so.
The second is retraction. If the source was wrong, or the situation changed, you need a way to say "this is no longer true" that does not require deleting the history. Tan's own description of the overnight cycle includes the agent fixing its own citations and consolidating memory. That is a maintenance loop, not a generation trick, and it is the difference between a notebook and a system.
There is a typed knowledge graph underneath, which matters for a specific kind of question. "Who works at this company" is not a similarity problem; it is a walk. GBrain's relationship retrieval lifts production R@5 from 0.6626 to 0.7241 on its own benchmark, with a graph-oracle upper bound at 0.9791. Note what that last number is: an upper bound, not a product score. Publishing both is the kind of thing that makes the rest of the claims easier to believe.
Installation, honestly
This is a Bun project, and the README is unusually direct about a trap. GBrain is not distributed on npm. The npm package named gbrain is unrelated and can shadow the real binary, so the instruction is to install only via bun install -g github:garrytan/gbrain#stable, or clone the stable branch and link it. If you installed the wrong one, gbrain doctor gives you the fix.
For a minimal local setup you can run gbrain init --pglite --no-embedding — keyless, no server, no Docker. To reach it from other machines, gbrain mcp expose publishes it over MCP on a Tailscale tailnet, with a --funnel flag for cloud agents. The claim is a thirty-minute install. My guess is that thirty minutes is accurate for someone who already has Bun and a Tailscale account, and optimistic for everyone else, which is normal.
Hooking it into an existing agent is deliberately unglamorous: MCP for the general case, a plugin marketplace entry for Claude Code and Codex, a documented path for hosted brains with OAuth. The system is designed to attach to the agent you already use, not to replace it with a new personal assistant identity.
The part that should worry you
A company brain and a personal brain are different threat models, and the README knows it. Remote authenticated clients are constrained by source grants, operation grants and visibility filters. Local files and shared database credentials sit on the other side of a trust boundary — source-level isolation does not protect callers who can read the disk.
That is a more honest paragraph than most projects manage. It also means the answer to "can we share this with the team" is not yes or no. It is a list of specific access paths that were tested, and a list of ones that were not.
There is a second, subtler risk that comes with any long-lived memory: it accumulates your mistakes. An agent that remembers everything remembers the wrong hypothesis you spent three weeks on, the vendor you dismissed for a bad reason, the person you mischaracterized in a note. Correction and withdrawal are the mechanisms that keep this from becoming a sediment layer of your worst judgments. Using them is a habit, not a feature, and most teams will not build the habit until something embarrassing surfaces.
What I would take from it
Not the page count. The page count is a flex, and an honest one, but it is not transferable.
The transferable idea is the ordering. Keyword retrieval first, model-free, working on day one. Semantic layers later, when you can measure whether they help. Graphs for the questions that are actually walks. Synthesis only where a cited answer beats a list of chunks. And a maintenance loop that runs whether or not anyone is watching, because a memory that is not maintained is just a pile.
There is also a smaller, sharper tool in the repo worth a look: gbrain skillopt, which treats a SKILL.md as a trainable parameter — generate a benchmark from the skill, strengthen the judges, let the optimizer propose edits, keep only the ones that measurably score higher. Twenty minutes, about a dollar of API calls. It is the same philosophy in miniature: if you cannot measure the improvement, you do not get to keep it.
Related articles
Kling's Technical Report Says the Bottleneck Is Not the Renderer
A video model team published an argument instead of an announcement, and the argument is that the renderer is the easy part.
The Three-Panel Film Strip Is the First AI Format With a House Style
Three frames, one person, cool tone, subtitle burned in. A generation trick turned into a distribution format.
Luma and Runway Moved the Model Into the Chat Box
Neither company shipped a better model this month. Both of them shipped a shorter path from question to render.
Disney Put an AI Lyric Video on Stage and a Lawsuit on File
Two things can be true at the same time, and the distinction the studio is drawing is about who authorized the training data.