A company brain is a single, current, structured representation of how a business actually works, its definitions, rules, and institutional knowledge, captured once and made readable by both people and AI tools. It is the place a company's meaning lives so that every tool and every new hire works from the same source instead of piecing it together from a dozen scattered places, and so no AI tool has to guess what the business means.
The term is new, and it is forming into a category right now. In its 2026 Requests for Startups, Y Combinator named "Company Brain" one of the ideas it most wants founders to build. That is a strong signal, and it points at a problem every company now shares: AI is being wired into how the business runs, and every AI tool hits the same wall. It does not know the company.
This piece defines the term, traces where it came from, walks through what separates a real company brain from the things commonly mistaken for one, breaks down the four jobs a working one has to do, shows how to start building one, and is honest about who needs one and who does not.
Why the idea is showing up now
For most of software history, a company's knowledge lived in people's heads, in wikis, and in chat threads. That was fine, because only people needed it, and people can ask around. If a new hire did not know how renewals were approved, they asked the person two desks over.
AI changed the requirement. When you point ChatGPT, Cursor, or your own agent at a question about your business, it cannot ask around. It answers in a single pass from whatever it can read in the moment. And the things that actually decide the right answer, how your company defines "active customer," which accounts are politically sensitive, the reason a specific customer gets a discount, were never written down anywhere a model can read them. So the tool does the only thing it can. It guesses, fluently and with confidence, and the guess is often wrong in a way nobody catches until it lands in front of a customer or in a board deck.
A company brain is the response to that. Write the company down once, in a form a machine can use, and stop every tool from guessing. The reason the idea is surfacing across the industry at the same time is that companies crossed a threshold: most are now running not one AI tool but several, and every one of them is failing at the same wall for the same reason.
A day this actually breaks
Here is what the abstract problem looks like on a Tuesday. A 90-person software company runs ChatGPT in the ops team, Cursor across engineering, a support assistant answering customers, and a finance lead who asks Claude about the numbers. Someone asks each of them, in effect, the same question: how are we doing on active customers this quarter.
The support assistant counts anyone who logged in this month. The sales assistant counts anyone under contract. The finance model counts anyone who paid an invoice. Three tools, three numbers, all defensible, none the same, and all of them built on a definition the company never actually agreed on in a place the tools could read. Nobody notices until two of the numbers show up in the same meeting. Then the entire AI program loses trust in an afternoon, because if the tools cannot agree on what a customer is, no one is going to let them near a forecast.
That is the problem a company brain exists to solve. One definition of "active customer," written down once, that all four tools read from and all four apply the same way.
Where the term came from
Nobody owns the phrase "company brain," and that is part of why it is worth taking seriously. The name converged from several directions inside about six months, which is what it looks like when a real problem starts forcing its way into language.
In December 2025, the venture firm Foundation Capital published an essay arguing that the next layer of enterprise value is what it called context graphs: a living record of how a company actually makes decisions, stitched across systems and time, so that precedent becomes something an agent can search. Different word, same wall. Agents cannot work inside a company whose reasoning was never written down.
In April 2026, the idea got a personal-scale proof, twice within a day. Andrej Karpathy published his LLM wiki idea: instead of having an AI re-retrieve raw fragments on every question, let it compile source material into a persistent, interlinked Markdown wiki that it maintains and corrects over time. A day later, Garry Tan open-sourced GBrain, the Markdown-based memory layer his own agents read from and write to. It crossed 5,000 GitHub stars in its first 24 hours, which tells you how many people were already trying to build the same thing for themselves.
Then Y Combinator, in its 2026 Requests for Startups, gave the company-scale version its name and asked founders to build it.
Notice who is in that list. A venture firm mapping decision traces, a researcher compiling wikis, a CEO publishing his agents' memory, and an accelerator naming the category. Four different seats, one conclusion: AI is only as useful as the context it can read, and almost nobody's context is written down.
One distinction worth keeping sharp, because it decides what you build. Karpathy's wiki and GBrain are personal brains. One person decides what is true, and that person is also the only reader who has to trust it. A company brain has the harder version of the problem: fifty people with three definitions of "active customer," rules that carry authority only if someone decided they should, and readers who are allowed to see different things. That difference, one mind versus an organization, is why the rest of this piece spends so much time on governance and trust rather than storage.
What a company brain has to do to be real
This is the part the term hides, and it is where most attempts fall short. Most things sold or built as a company brain are one of a few older shapes, and each solves a different problem than the one AI actually has.
The pattern across the first four is the same. They were built for a person to read, not for a machine to resolve against while it answers. A company brain is only as useful as whether your AI can read it at the moment it needs it. A brain you browse is a wiki. A brain your AI resolves against while it works, governed and traceable, is a different thing, and the architecture term for that thing is a context layer.
The four jobs a working company brain does
Strip away the metaphor and a real company brain comes down to four jobs. The order matters, because skipping any one is where these projects fail.
It captures the knowledge, including the unwritten kind. The easy half is the definitions and metrics that already live in dbt and the warehouse. The hard half is the knowledge that lives only in people's heads: how renewals actually get approved, why an account is sensitive, the reason a customer got a discount three years ago. A real company brain gives the people who hold that knowledge a fast way to write it down, with an owner and a last-validated date attached, so it becomes a fact the company keeps rather than a memory that leaves when the person does. This is the surface most "company brain" attempts skip entirely, and it is the one that separates a brain from a glossary.
It governs what becomes canonical. Not everything someone writes down should carry authority. If every note is treated as truth, the brain becomes a dumping ground and the AI cites contradictions with equal confidence. A working company brain has a review step: someone decides what is canonical, what stays draft, and what gets retired. This is the difference between governed knowledge and a shared folder.
It serves the context to AI at the moment of the answer. A brain that lives in a document your AI cannot query at inference is just storage. The knowledge has to reach the tool while it works, filtered by who is asking and what they are allowed to see, so a support assistant and a finance agent each get what they should and nothing they should not. The open standard for this is MCP, which lets ChatGPT, Cursor, Claude, or your own agents all read from the same source. This is the job that turns a document into AI that knows your company.
It makes every answer traceable. The last job is what makes the brain safe to rely on. When a tool answers, you should be able to trace the answer back to the exact rule and definition it used. That way a wrong answer points you to the source that needs correcting, and you can tell the difference between the model failing and the rule being out of date. Without this you cannot put AI in front of a customer, because you cannot explain what it knew.
Do these four and you have not built a document. You have built a context layer, which is the honest architecture name for a company brain your AI can actually use.
The failure mode nobody warns you about
The most common way a company brain goes wrong is not that it is never built. It is that it is built as a documentation project and then rots. A team writes everything into a wiki over two weeks, points AI at it, and calls it done. A month later a policy changes, nobody updates the page, and the AI keeps citing the old policy with full confidence, because the brain told it to.
A stale company brain is worse than none, because it is wrong with authority. A person who finds an out-of-date wiki page shrugs and asks a colleague. An agent quotes it to a customer. This is why the build approach matters as much as the content: a company brain has to be seeded from the tools you already run and kept current from those sources, not hand-authored and left to decay.
How do you build one
The short version, for orientation. The full walkthrough, with the exact starter notes and the config to copy, is in How to Build a Company Brain.
- Start where the disagreement is. Pick the ten terms and rules that get two different answers in the same meeting. At most companies that list starts with "active customer," "qualified lead," churn, and the discount policy.
- Write one note per fact. Each definition or rule stated in one sentence a machine can quote, with the known conflicts named instead of smoothed over.
- Attach an owner and a date to every entry. A definition without those is a rumor. This is also what makes staleness visible later, instead of silent.
- Serve it to your AI over MCP. The brain only counts if your tools read it at the moment they answer, not if it sits in a folder.
- Review on a schedule. A weekly pass over what changed, so drift gets caught as a prompt to confirm rather than discovered in front of a customer.
You can genuinely start this in Obsidian over a weekend, and the guide shows how. The same guide is honest about where the DIY version stops: roughly at the moment more than one person needs to trust it.
Who actually needs one
A company brain is not for everyone, and the honest version of this page says so.
It pays off most for the company in roughly the 40 to 200 person range that has already deployed several AI tools and watched them get the business wrong. That company is large enough that the founder's head is no longer the single source of truth, and small enough that a quarter-long enterprise governance rollout is a non-starter. The pain is visible to leadership, usually because a tool gave a confidently wrong answer that someone important saw.
Below that size, the founder still is the company brain, and adding a system for it is overhead you do not need yet. Above it, in the large enterprise, dedicated data governance teams and catalog platforms are already in the building, and the problem looks different. If you are in the middle, running three or more AI tools without a platform team to babysit them, a company brain is the thing that makes all of them reliable at once. And if the tools you care most about are agents acting on your behalf rather than people asking questions, Company Knowledge for AI Agents goes deeper on that case.
The point
A company brain is a good name for a real shift: the company's meaning and rules moving out of people's heads and scattered tools into a single source that both people and AI can use. The part the name hides is that the source only matters if your AI can read it at the moment it answers, governed and traceable. That is the work, and it is what we are building Sento to do.
Whether the category ends up called a company brain, a context layer, or something else, the outcome is the same, and it is simple to state and hard to fake: AI that knows your company. Your AI doesn't know your company. Sento fixes that.
What is a company brain in simple terms?**
A single, current source of how your business works, its definitions, rules, and institutional knowledge, written down once and made readable by people and AI tools alike, so nothing has to be pieced together from scattered places and no AI tool has to guess.
Isn't a company brain just a wiki or a chatbot over our docs?
That is the most common version, and it is the one that falls short for AI. A wiki and a doc chatbot return text written for a person to read, with no authority and no way to filter by who is asking, and they are not queryable by an agent at the moment it answers. A working company brain serves governed context to AI tools at inference, traceable to source. A wiki is a good place to seed it, not a substitute.
How is a company brain different from enterprise search like Glean?
Enterprise search finds documents for a person. A company brain serves the company's actual definitions and rules to AI tools while they work, so the tool applies the rule rather than handing a human a list of links. Different consumer, different moment of use.
How is it different from a data catalog like Atlan?
A data catalog documents metadata for analysts to browse and for BI governance. It describes your data. A company brain serves the meaning and rules to AI tools at inference and traces what each answer used. Catalogs are built for people browsing; a company brain is built for machines resolving.
Where did the term "company brain" come from?
It converged from several directions between late 2025 and mid 2026: Foundation Capital's context graphs essay (December 2025), Andrej Karpathy's LLM wiki idea and Garry Tan's open-source GBrain (both April 2026), and Y Combinator naming "Company Brain" in its 2026 Requests for Startups. Different names from different seats, all describing the same wall: AI can only be as useful as the context it can read.
How is a company brain different from a second brain?
A second brain is personal: your notes, your truth, and you as the only reader who has to trust it. Tools like Obsidian and GBrain work well at that scale because one person decides what is canonical. A company brain is the organizational version of the problem: many authors, conflicting definitions, rules that need authority behind them, and readers who are allowed to see different things. Storage is the easy part. Governance is the difference.
Why did Y Combinator name "company brain" a top idea for 2026?
Because every company is deploying AI and every tool hits the same wall: it does not know the business. The knowledge that runs a company was never written down in a form AI can read. YC sees building that layer as one of the most important problems in software right now.
What makes a company brain actually work with AI?
Four things: the knowledge is captured with an owner and a date, governed so only canonical context carries authority, served to any AI tool at inference over MCP, and traceable so every answer points back to the rule it used. The architecture term for this is a context layer.
How do you build a company brain without it going stale?
Seed it from the tools you already run rather than authoring from a blank page, so most of it updates automatically, and surface drift as a prompt to confirm rather than a chore to remember. Every entry carries an owner and a last-validated date, so anything aging gets flagged for review instead of silently rotting.
