What the hell does a “company brain” even mean?
August 24, 2026
Say “AI for your organization” in 2026 and within a sentence somebody reaches for the phrase. A company brain. An org brain. A second brain for your business. Y Combinator’s latest request-for-startups lists “Company Brain” as a category they want built. The big vendors sell under some version of it. People nod when they hear it.
I’ve started asking the people nodding what they think it means. I recommend trying this. You get intelligence, vaguely. Memory, vaguely. Something about “it knows your company.” What you never get is an answer precise enough to build from. That gap is worth noticing, because thousands of teams are building from it right now.
I can tell you what it means, because I’ve spent the last year building one, and the build forced me to find out. Not in the abstract: I maintain roughly eighty rules whose only job is stopping a language model from inventing things about my own company, my extraction prompt is on version fourteen, and about forty-five of my database tables exist purely to keep one catalogue honest. You learn what a “company brain” is when every wrong answer files a bug against your definition of it. And the answer was sitting in the one working brain we all carry around. Yours is not one thing. It’s four, and each part does a job the others can’t:
An org brain, if the phrase is going to mean anything, needs the same four parts: something that reasons, something that keeps every record organized and hands the right ones over at the right moment, something that notices an event and acts without being asked, and somewhere for the lessons to go.
The unbundling matters because the four parts come to you in four different ways.
The intelligence you rent. The librarian you build. The reflexes you wire. And the rules you earn — every one of them costs you a wrong answer first.
The intelligence is Claude, GPT, Gemini: trained on the written output of the world, a few dollars per million tokens, identical for every company on earth. The librarian is built from your own records, and nobody can sell it to you because nobody sells your history. The reflexes get wired one trigger at a time. And the rules accumulate the expensive way.
Which puts the industry in a funny position: the vendors selling you a “company brain” are, mostly, selling you the intelligence. The one part you already own.
Sketch is my bet on the other three. I’ll be honest that it was the slow way to build. While the market shipped agent demos, I spent my months on entity resolution and record-keeping, the least glamorous work in AI. Early on, my own pipeline minted a company called “Gmail” because fifty-seven people in the data had gmail addresses. It once decided a market-intelligence client had 203 products when they had three. This note is about why I’d make the same bet again, and about the week the bet showed itself more clearly than I could have staged it.
In the third week of August 2026, I spent a few days deciding whether to take Sketch’s repository private. Not a small call: six months of work sat in that repo, and flipping the switch would erase 179 stars and orphan 34 forks overnight. It felt heavy enough that I thought it through with Fable, Claude’s top model, over one long conversation. And somewhere in that conversation, the parts of the “brain” came apart in front of me, cleanly, like a demonstration I couldn’t have scripted.
The first job Fable did looked like wisdom. It walked me through the trade-offs of source-available licenses, what security auditors ask for in practice, when open source works as distribution and when it just leaks your thinking to competitors. Good counsel, and none of it came from me. That’s the rented intelligence: Fable has read enough of the world to know how these things go. Impressive, and completely unremarkable. Every founder on earth can rent the same advisor.
The second job is the one worth slowing down for. Midway through, I needed the full history with one prospect: a company that had tried our hosted product, then forked our open-source repo to run it themselves. Their engineer had an open pull request we’d never merged. And ten days earlier, in a WhatsApp group, I had told him we’d merge it “next week.”
Fable pulled all of that together in under a minute, by sitting down at Sketch’s reading desk over MCP and asking my librarian. The first demo call in March. The meeting in August where their goals changed. The exact WhatsApp message with my half-promise in it. The contents of the PR. (The decision on the table was about Sketch’s own future, and Sketch was the one arming it, the product making its case in the middle of a conversation about it.) And with the records on the table, Fable’s advice changed character. It stopped being general wisdom about open versus closed source and became: you made this person a commitment, and your message needs to address it, or it will read like a dodge.
Same Fable, same conversation. The first job ran on knowledge of the world. The second ran on knowledge of us. Fable didn’t get smarter between the two moments. It got informed, by the part of the brain I had to build, because nobody would rent it to me.
Once I had the unbundling clear, I went looking for places it would break. It kept not breaking, in places that surprised me.
History got here first. In 1935, Paul Otlet described his life’s work as a “mechanical, collective brain” for the world’s knowledge. What he built, in Brussels, was the Mundaneum: twelve million index cards cross-referencing the world’s books, answering around 1,500 queries a year by mail. The grandest “brain” of the pre-computer era was a card catalogue with clerks. (The Nazis destroyed 63 tons of it in 1940 to make room for an art exhibition, which tells you something about how institutional memory dies.) And the mislabeling never stopped: today’s “Build a Second Brain” movement, the biggest personal-knowledge brand of the decade, teaches capturing and filing notes into folders. A filing system, sold as a brain.
So do your best teams. Daniel Wegner’s research on transactive memory found that good teams don’t all remember everything. They maintain a shared sense of who knows what, and retrieve through each other. A well-run team is a directory plus retrieval plus a colleague who taps your shoulder when something needs you. Plus, notice, a rulebook: “never ship on Fridays,” “always loop in legal on pricing.” Every mature team runs on laws it wrote after something went wrong. Nobody on it shares a brain, and it works precisely because nobody has to.
Even the memory startups. This is the one that makes me laugh. Look inside the products marketed as AI memory today, such as Zep or Mem0, and you find knowledge graphs and record-filing systems. None of them touch the model’s weights. The memory industry, at the implementation level, is a library industry. They know. The branding just pays better.
This would be pedantry if the words led to the same roadmap. They don’t. I know because I’ve paid for stretches of the wrong one myself.
Teams that set out to build a brain reach for cognition. They fine-tune models on company data. They build memory that “learns over time.” On the first idea the evidence is now damning: Stanford tested commercial fine-tuning on exactly this job and found models absorbed new facts at roughly 37% accuracy, and updates to existing facts at 19%. The practitioners’ rule of thumb has settled into “fine-tuning is for form, not facts.” The other shortcut, stuffing everything into the model’s context window, degrades too. Accuracy falls off well before the advertised limits.
Building the librarian is much less glamorous, and I can tell you exactly what it is, because it’s where my year went. A catalogue: one entry per real thing, per person, per client, per project, no matter how many tools it shows up in. When “NP,” “Northwind Partners,” and a bare WhatsApp number are the same relationship, the catalogue has to say so. Getting that wrong is not hypothetical. I’ve had a stale contact record quietly soak up 129 files that belonged to someone else. My matcher is explicitly forbidden from trusting reordered names, because “Li Wang” and “Wang Li” can be two different people, and a catalogue that merges them has poisoned every answer about either. This is entity resolution. It’s the hard part, hard enough that I wrote a separate note about it. The rest is well-organized records, chunked so a question pulls up the relevant passage rather than the whole archive, and retrieval in context: the right records handed to the model at the moment it’s reasoning about you.
A library is also inspectable, which I consider a feature you should refuse to live without. When it serves up a wrong record you can see which one, fix it, and that mistake is gone for good. When a fine-tuned brain gets something wrong, your only lever is another training run.
The obvious objection: a librarian sounds passive, and what buyers want in 2026 is not a better archive. Fair. This is exactly why the third part exists, and why it’s a part of the product and not a feature of the model.
A meeting ends. Sketch reads the transcript, works out what was promised and by whom, and routes the follow-ups to the people who owe them, on Slack or WhatsApp, before anyone asks. Nobody prompts this. It’s a reflex: something happened, the system noticed, work moved. The same pattern runs underneath everything interesting Sketch does. A promise made in a chat gets tracked, a project that changed state gets flagged, a review lands with the person whose call it is.
Nature figured out the design principle long ago: reflexes don’t route through the expensive brain. My trigger layer runs deterministic and cheap, the way your blink does. The genius is not consulted about every ball flying at every face. Sketch files, routes, and notices on cheap intelligence, because shelving doesn’t need a genius. Then, when a question deserves the best brain money can rent, you plug that brain into the same catalogue over MCP and let it read. That’s what the repo decision was: cheap reflexes and filing all year, Fable at the reading desk for the one call that mattered. Genius is for reading, not shelving.
Now the fourth part, and the one I’d have argued against a year ago. I told you nothing in the librarian’s job description learns. That’s true, and it raises the obvious question: my system gets visibly better every month. Where does the improvement live, if not in a learning model?
In the rulebook. Remember the “Gmail” company my pipeline minted? The fix wasn’t a smarter model. It was a rule: personal email domains never vouch for a company’s existence. The stale contact that soaked up 129 files became a rule about when a match must abstain. Eighty-odd times now, the expensive rented brain (or I) made a judgment call once, and instead of hoping we’d make it again next time, the judgment got compiled into law: deterministic, cheap, permanent. Exactly what your basal ganglia do with clutch-brake-gear. The lesson stops being thought and becomes reflex-shaped, except this one is earned.
This is how a stateless brain gets a memory of its mistakes. You can’t fine-tune the lessons in; the numbers above say that fails two times out of three. But you can legislate them. And legislated learning has a property weight-learning will never have: you can read it. Every lesson in my system is a rule with a name, a reason, and the incident that created it. When a rule turns out wrong, I repeal that one rule. When a fine-tuned model learns something wrong, you retrain and pray.
Tanush, one of my engineers, read a draft of this and did something useful with it: he mapped the four faculties onto the four things we’ve built. It’s the moment the metaphor stopped being a metaphor for me.
Intelligence is the agent: the rented model, doing the reasoning. The librarian is the context graph: the catalogue of every person, client and project, resolved across every tool. Reflexes are automations: meeting ends, follow-ups route, nobody asked. And rules are skills, which is a better mapping than the one I’d written, because a skill genuinely is a compiled procedure. Somebody worked out the right way to do a thing once, wrote it down, and now it gets followed without being re-derived. That’s the whole definition of a habit, sitting in a file.
Four faculties, four surfaces, all four working. Which brings me to the part that isn’t.
Today, rules enter my system through me. Something goes wrong, I understand why, I write the gate. Eighty times.
The system writes only its smallest laws itself: reject a bad match once and it’s never proposed again, let a candidate get skipped three runs straight and it retires on its own. Real, but small: those are laws about the catalogue’s own hygiene, not about how the company works.
The missing piece is the loop that would close it: intelligence reviewing the record on a schedule, noticing that the same judgment has been made fourteen times, and proposing the rule itself.
The brain has a name for this. It’s consolidation, and it happens while you sleep. The hippocampus replays the day to the cortex, and whatever recurred gets compiled into general knowledge and routine. Notice the shape of it: periodic, offline, and selective. Your brain doesn’t consolidate everything that happened, only what repeated, or what mattered. It runs a nightly pass over the index and promotes the patterns into law.
That’s the machine I need to build, and I’m fairly sure I already own most of the parts. My weekly pass is already a periodic agent that reads the record and proposes; the spine underneath it is already propose, validate, human-approve, apply through a ledger that can undo. Entities were the first object that spine curated. Rules would be the second. The lesson to distill is sitting in the record already — every time a human corrected the machine, that was a labelled example of a rule that didn’t exist yet.
I want to be precise about what I’m claiming, because this is exactly where a company brain pitch usually starts lying. The four faculties are built and running. The consolidation loop is not. It’s the next thing I’m building, and the reason I’m confident about it isn’t that it’s easy. It’s that the hard part was never the intelligence. It’s having a record clean enough that a machine reading it would draw the right conclusion. That part took a year.
When it does run, it will still ask before it legislates. The model proposes, deterministic code validates, a human approves, and the whole thing can be revoked. The system won’t learn in the dark. It will draft legislation and ask.
That’s the shape of a company brain, and none of it required building intelligence: a rented mind, a built library, wired reflexes, and a rulebook (for now, written by hand). The industry keeps trying to sell the first part. The other three are the product, and the loop between them is the thing worth building next.
The most valuable knowledge in a company is tacit, and it never makes it into any record. This is the objection that killed 1990s knowledge management, and it’s half right. A library captures the traces judgment leaves behind: who was in the room, what was said, what was promised. It does not capture the judgment itself, and I’d rather admit that than overclaim. But two things have changed since the SharePoint era. Those systems died because keeping them current required human filing effort that no organization ever sustained, and that constraint is simply gone: the librarian now reads everything itself, so the catalogue stops rotting when people stop filing. And notice who this objection wounds more. A product that promises to think like your company is claiming to hold exactly the knowledge that can’t be captured. A library never claims to hold more than it does. I’ll take the honest architecture.
If the value in this category were the intelligence, the model labs would own it outright. They own the intelligence. But every company rents the same minds at the same price, which means intelligence is the one layer of this stack that cannot differentiate anyone. I find it strange how many pitch decks haven’t absorbed this.
The other three parts are different for a stubborn reason: they’re made of you. The catalogue is built from your records (your calls, your threads, your promises) and it can’t be bought, because nobody sells your history. The reflexes are wired to your operating rhythm. And the rules are laws derived from your own mistakes and your own evidence: a competitor can copy the mechanism, but they cannot derive your laws without living your year. All three compound. Every week the catalogue gets more complete, and every wrong answer makes the rulebook longer. The tools will keep getting cheaper. The shelves, the wiring, and the law stay yours.
So my bet, stated plainly: the companies that win with AI won’t be the ones with the smartest system, because everyone’s system is equally smart now. They’ll be the ones whose records were kept well enough, whose reflexes were wired tightly enough, and whose lessons were written into law, so that the rented mind, the moment it sat down, knew exactly what it was talking about, what had just changed, and what never to do again.
Otlet at least had an excuse. In 1935 a brain was the only machine for knowledge anyone had ever seen. Ninety years later the industry is still pointing at twelve million index cards and calling them a mind.
I built the librarian, the reflexes, and the rulebook. The brain, I rent — same as you.
We build Sketch. If you want to try it, talk to us →