How the company brain actually works.

This post opens the company brain and shows you the gears. If you are going to trust software with everything your company knows, you should be able to see exactly how it works, down to the shape of a single fact and what the system does when two of them contradict each other. No diagrams hiding the hard parts. We have argued elsewhere that your company's memory should not live in a wiki. This is the part where we stop arguing and show the machine.

The company brain is a living, sourced model of how your company runs. It builds itself from real activity, it puts a source on every fact, and it belongs to you. Each of those is a decision we made on purpose, and each one falls apart if you wave at it instead of showing it. So we will show it.

It builds itself, because the alternative never works

The brain has no data-entry step. No onboarding wizard asks you to fill in your company. The moment a system depends on a human to keep it current, it starts going stale, and that is not a discipline problem. It is structural. Nobody stops doing their job to go update a model of their job.

So the brain reads the work that is already happening: your email, your messages, your CRM stage changes, your documents, your calendar. When a deal moves from Proposal to Negotiation, when a thread in a channel reaches a decision, when a contract gets signed in Drive, the brain catches the event and extracts the facts and the edges between them. This runs on the activity stream, not on a clock. A deal moving stages produces new facts within seconds of the move, not on the next nightly batch. The same work that would have made a wiki stale is the work that keeps this current.

What one fact actually looks like

This is the part most AI memory skips, and it is the part that separates a model you can act on from a folder of plausible guesses. In the brain, a fact is never a bare string. It is a claim with structure attached. Here is a real one, the kind the brain stores when it notices an account going quiet:

{ "claim": "Acme Corp engagement is cooling", "subject": "company:acme-corp", "provenance": "inferred", "confidence": 0.62, "sources": ["email:thread-8841", "crm:acme/last_contact"], "last_confirmed": "2026-05-29", "pinned": false }

You can click any single thing the brain believes and see that whole object: where it came from, how sure it is, when it last held true, and what kind of knowledge it is. No belief is anonymous. The source is what makes it auditable. If the brain tells you a renewal is at risk, you do not have to decide whether to trust a black box. You follow the source ids back to the actual message and the actual CRM field and judge them the way you would judge any source. A fact without a source is a rumor with good production values, and we do not store those.

The provenance field grades what kind of knowledge a fact is, on four levels:

• →Observed: the brain saw it directly in the activity. The deal moved to closed-won, the message was sent, the document was signed. Firmest ground.
• →Derived: computed from observed facts by a rule you can inspect. Total pipeline value, days since last contact, the renewal date sixty days out. True as long as the inputs are true, and you can check the inputs.
• →Inferred: a reasonable conclusion the brain drew that was never stated outright, like the Acme fact above. This account looks like it is cooling; this person seems to own this process. Useful, and the softest of the four, which is exactly why it is labeled 0.62 and not presented as gospel.
• →Pinned: a fact you set by hand. This one overrides everything else, and we will come back to why it exists.

Freshness gets its own field because a company will not hold still. A perfectly true fact from eight months ago can be a lie today. The pricing changed, the owner left, the process evolved. The last_confirmed timestamp lets the brain tell the difference between something it verified this morning and something it has not seen reaffirmed since Q1, and treat the old one with the suspicion it has earned.

When sources disagree, you pin the source of record

Real companies contradict themselves constantly. The contract says one renewal date, the email thread implies another. Sales calls the deal six figures, finance has it at five. Two people each believe they own the same process. A naive memory system either picks one at random and hides the conflict, or averages them into mush. Both destroy the one thing you needed, which is the ability to know what is actually true.

The brain does neither. When two sources disagree, it surfaces the conflict instead of swallowing it. It shows you both claims, both sources, and lets you do the thing only a human can do: decide which one is binding. You pin it. The pinned fact flips to provenance "pinned" and wins from then on, while the losing claim stays visible in the conflict trail, marked superseded. The wrong source does not get a vote anymore, but you can always see that it once cast one.

This is the same approval-first instinct that runs through everything we build. The AI does the legwork of finding the contradiction. The human makes the call about which version of reality is binding. We are not interested in software that quietly resolves your company's disagreements on its own. That is precisely the kind of decision that should wait for a person.

Three facts, not four hundred

A search box can recall everything, and that is its problem, not its feature. Ask a company brain a question and get four hundred matching facts back, and it has handed the work straight back to you, the same way a dashboard hands you a chart and calls it insight. Recall is easy. Knowing which three facts matter to the question you asked is the hard part, and it is the only part worth having.

So here is an actual question, run against an actual graph. An AI Employee on the account asks the brain: why is Acme cooling? The brain holds roughly four hundred facts that touch Acme: every email, every stage change, every contact, every line item going back two years. It returns three, in this order:

• →Acme's champion, Dana, left in April (observed, confidence 0.98, confirmed from a bounced email and a LinkedIn change).
• →The replacement contact has opened zero of the last five emails (derived, confidence 0.91, confirmed this week).
• →Acme's renewal sits 47 days out with no meeting booked (derived, confidence 0.95, confirmed today).

It leaves out the other ~397, including a high-confidence observed fact that Acme loved the Q3 onboarding, because that fact, however solid, does not bear on cooling. Ranking is relevance first, then confidence and freshness as tiebreakers. The genuinely hard case is when those two conflict: a rock-solid observed fact from last spring against a shakier inference from this week. The brain does not blindly prefer the recent one. It weighs how much the question is asking about right now. For "why is Acme cooling," a present-tense question, the fresh inference about ignored emails outranks the stale observed fact about a beloved onboarding. For "what has Acme historically valued," the ranking flips. Relevance is judged against the question, not the calendar.

And because the model is a graph, the answer is not three loose facts sitting next to each other. The edges connect them. Dana left, which is why the replacement is disengaged, which is why the renewal has no meeting. The brain can walk that path from the symptom you noticed to the cause three months back. That walk is the thing forty disconnected tools could never do, because each of them only ever held one node and none of them held the edges.

What the brain is actually made of

Underneath, the model is a graph: entities for nodes, relationships for edges. The entities are the nouns your company is built from. The brain tracks, among others:

• →People: who knows what, who works with whom, who to ask about a given thing.
• →Decisions: what was decided, when, by whom, and the reasoning that did not survive in anyone's memory.
• →Processes: how work flows, read from the activity itself rather than from a doc that describes an aspiration nobody follows.
• →Customers and deals: the accounts, the opportunities, the stage they are in, and the history that explains why.
• →Money: the numbers attached to the deals and the company, each carrying the source and freshness that tell you whether to trust the figure.
• →Intent: where things are heading. The goals, the bets, the direction the activity is pointing at.

The edges are what an org chart and a CRM both miss. Knowing a person and knowing a decision is filing. Knowing that this person made that decision, which shaped this process, which is why that customer churned, is the part that lets the Acme walk happen. The nodes are cheap. The edges are the brain.

It is yours, and you can take it with you

The brain is built on Platos, Winsen's open agent runtime. Platos is open source and on GitHub, so the substrate your company's memory lives on is something you can read, fork, and run yourself. Your brain is isolated to your company. It never trains a shared model. And there is an export button: the whole graph comes out as JSON, every node and every edge with its sources and timestamps intact, plus a SQL dump for the underlying tables. It opens in anything that reads JSON. No re-import fee, no proprietary format, no lock-in clause to negotiate your way out of.

Think about what the alternative asks you to hand over: everything your company knows, locked inside a vendor's format, leaveable only by abandoning it. That is not a tool, it is a hostage situation with a nice UI. Because your brain is yours and exportable, the longer you use it the more leverage you have to leave, not us. That asymmetry is what keeps us honest. We have to earn the next month, every month, because you could walk out with the file at any point and we know it.

That is the whole machine. It builds itself off your activity, so it stays current. Every fact carries a source, a confidence, a freshness, and a provenance level you can click. Conflicts surface for you to pin instead of getting hidden. It returns the three facts that answer the question and walks the edges between them. And you can leave with all of it, in a file format you already use, today. That last one is the part to remember, because you cannot trust a box you are not allowed to open, and this one opens.