A company brain compared with what you already have
Nobody starts from nothing. By the time a team goes looking for a company brain they are already paying for a wiki, a search tool that indexes the drive, and quite possibly a platform for building agents. The reasonable question is not whether the idea is good. It is whether any of the three things already on the invoice does it.
Three comparisons, each fair to the thing it compares against, because two of them are genuinely good at what they were built for.
| Approach | What it gives you | Where it breaks | Right when |
|---|---|---|---|
| Wiki | What people deliberately wrote down, with intent and nuance | Nobody updates it, and everyone keeps trusting it anyway | The knowledge is judgment a system could not have derived |
| Retrieval | The passages that match the question | Fragments contradict each other, carry no date, and cost tokens every time | You need to find a specific document |
| Agent platform | Orchestration, tools, somewhere to run the work | Ten agents each taught the company separately, drifting from week one | The context question is already answered |
| Company brain | One current account, reconciled, sourced and dated | Inherits whatever the source systems get wrong | Several people or several agents need the same answer |
Is a company brain just a wiki?
No. A wiki holds what someone chose to write down; a company brain holds what the systems already record. That is why the two age so differently.
A wiki is a place people write. That is its definition and its whole strength: intent, judgment and nuance that no system could have derived, put down deliberately by someone who understood it.
Its failure mode is equally definitional. A wiki is only as current as the last person who felt like updating it, and updating it is permanent work with no deadline, which makes it everyone's fourth priority forever. So the page that describes your onboarding is accurate for about a quarter, then becomes a confident record of how things used to work. Nobody deletes it, because nobody is sure what has changed. Everybody stops trusting it, and it stays in the search results anyway.
A company brain is not a better wiki. It is written from the tools instead of by hand, which means rerunning it is cheap and it can be current by default rather than by discipline — a single source of truth that holds because nobody has to maintain it, rather than because everybody agreed to. The tradeoff is real and worth stating: you lose the deliberate essay a thoughtful person would have written. The right arrangement keeps both, and expects different things of them.
Is a company brain just RAG?
No: retrieval returns the passages that match a question, and a company brain returns a position somebody already reconciled from them, with its date attached.
The standard technical answer is to index everything and search it: chunk the drive, embed it, and let the assistant retrieve passages at question time. It has been the default since the 2020 paper that named it, it works, and for finding a specific document it is the correct tool.
The gap shows up in what comes back. Retrieval returns passages, ranked by similarity, from documents that may contradict each other and usually carry no date. The assistant is then asked to reconcile four fragments written by three people across two years, and it will do it — confidently, with no signal about which one is still true. Ask ten people the same question and you get ten answers assembled from ten different fragment sets.
A company brain has already done the reconciling. Somebody, or something, has read across those sources and written down the current position, with its source and its date attached, before the question was ever asked. That is the difference between a search box and an account: one hands you the evidence, the other hands you the conclusion and shows its work.
It is also cheaper to work with, which matters more than it sounds. A distilled page is a fraction of the tokens of the raw passages it was drawn from, and that difference is paid on every single question.
The gap is understood well enough that the retrieval field is closing it from its own side: GraphRAG and Graphiti both add a structured layer over the chunks precisely because ranked passages were not enough. That is the same conclusion from the other direction.
Do agent platforms already do this?
Not the same job. Agent platforms are the loud half of this market, and the work they do is real: orchestration, tools, a place to build and run things. What they do not supply is the thing the agents read from.
But an agent is only as good as what it knows when it starts. Point ten agents at a company with no shared account of itself and you get ten agents each taught the business separately, by different people, at different times, drifting apart from the first week. The platform question and the context question look like the same question and are not. Context is underneath.
That is where this article stops, because the named comparisons have their own pages: Tulina vs. Dust puts it against an agent platform directly, and Claude with Tulina pairs it with an assistant that already has connectors. Both are more specific than anything that belongs here.
When you do not need one, and how to choose
The honest answer: if the company is small enough that everyone already knows everything, you do not have this problem yet. Two founders and an engineer share context by sitting in the same room, and any system for holding it is overhead against a problem they do not have.
It starts mattering at the point where somebody has to ask someone else how something works, and it matters a lot once AI is doing enough of the work that nobody is checking every answer. That threshold arrives earlier than it used to, which is most of why the term exists at all.
The choice between the four is less a ranking than a question of which failure you are actually hitting. If nobody writes things down, a wiki is not the answer and neither is anything else until that changes. If the documents exist but nobody can find them, retrieval is the correct tool. If the answers exist but disagree with each other and carry no dates, that is the reconciling problem, and it is the one a company brain is for.
If you would rather build than buy, several of the parts are open source and genuinely good; the route there is how to build a company brain, which covers the mechanism and surveys the projects. If the premise is still what you are weighing rather than the options, why your AI does not know your company is the argument underneath all three comparisons.
Common questions
- Do we have to replace our wiki?
- No. A wiki holds intent and judgment that no system could have derived, and that is worth keeping. The right arrangement runs both and expects different things of each: deliberate writing in one, the reconciled and dated account in the other.
- Is a company brain just RAG with better marketing?
- They answer different questions. Retrieval finds passages that match a query and leaves you to reconcile them, usually with no dates attached. A company brain has already done the reconciling and written down the current position with its source and date, which is why it also costs a fraction of the tokens on every question.
- We already have an agent platform. Is that not the same thing?
- An agent platform gives you orchestration, tools and somewhere to run the work. It does not give the agents a shared account of the company, so ten agents get taught the business separately and drift apart from the first week. Context sits underneath the platform question rather than beside it.