A company brain is a shared memory of what your organisation knows but never wrote down: what was decided on a call, why it was decided in a Slack thread, what a client mentioned in passing. It keeps learning from everyday work and serves that knowledge to people and AI agents, each seeing only what they are allowed to see.
Our newsletter
Get tutorials, AI agent recipes, webinars, and early product updates in your inbox every two weeks
Why is everyone talking about it now?
Y Combinator put "Company Brain" on its Summer 2026 list of startup ideas it wants to fund. The reasoning, as reported, is that the biggest blocker to AI automation is no longer the model. It's domain knowledge, which is "trapped in Slack threads, email chains, Notion docs, Jira tickets, and people's heads."
Meta's analytics team published how they built an internal "AI second brain" that reads meeting notes, tracks projects and builds on earlier conversations. It went from zero to more than 63,000 installs in three months, with 10,000 daily users across engineering, legal, finance, sales and other teams (Analytics at Meta).
And the problem it goes after is old. A 2018 Panopto study found that 42% of institutional knowledge is unique to the individual who holds it (Panopto). When that person changes role or leaves, the knowledge goes with them.
Isn't that just a context layer for agents?
Most AI projects I see in large organisations start with a context layer: give the agents access to data that already lives in databases. The CRM, the transaction history, the policy documents, the data warehouse. That's useful, and it's the right place to start. But that data was already recorded, structured and governed. Someone decided it was worth keeping.
My take is that a company brain is about the other part: the institutional (or tribal) knowledge that was never recorded or structured. In a bank, for example, it's why a client was given an exception on a fee, or what the risk committee was really worried about. Or the moment on a call when a client mentioned they were thinking of moving part of their portfolio. None of that is in a table. It's in transcripts, chat threads and people's heads.
When your agents automatically remember last week's meeting and the Slack thread that followed it, without anyone writing a summary, it starts feeling like a superpower. It brings context that, until recently, nobody even considered possible to make available to an agent.
Why do most teams leave the best context out?
Because it's private. The conversations that hold the most valuable knowledge are also the ones with client data, deal details, personnel issues and everything else you can't show to the whole company (or to every agent).
So the usual answer is to not ingest them. Teams connect the public channels and the shared drive, and leave out the private channels and meeting transcripts. The result is a company brain that only knows what was already public, which is mostly what was already recorded somewhere. That's a context layer again.
For regulated industries like financial services there's a second requirement: being able to show who could see what, and when. Employees also have to trust it, and they won't if they can't tell who will see what they say on a call.
This is why I think permissions are the most important part of a company brain. If access is enforced properly, you can ingest everything, because each memory stays visible only to the people who should see it. Without that, the safe choice is to leave the best context out.
Where does SurrealDB fit?
There are two layers, and we build both.
The database
SurrealDB keeps documents, graph relationships, vector embeddings and full-text search in one engine, and queries them together. That matters for a company brain because its knowledge is connected: people, clients, meetings, decisions and the teams they belong to. An agent can search by meaning, follow the relationships, and filter by date in a single query.
Permissions are defined in the database, on each record, and they're checked inside the search itself. Checking them after the search, in the application, is the easy mistake: one forgotten filter and an agent sees everyone's results. Access can also follow relationships. If access to a memory depends on team membership, it changes the moment someone joins or leaves the team, with no list of permissions to copy and keep in sync.
Agent Memory
SurrealDB Agent Memory is the memory layer we're building on top of SurrealDB. It ingests conversations and facts, and lets agents recall them. Four ideas decide who sees what:
Contexts: separate memories with nothing crossing between them (for example, one per legal entity or business unit).
Scopes: every memory is tagged with who it belongs to. A call between two people can be scoped to exactly those two, and a team's memories to that team.
Grants: what each person, or each agent acting for them, may read or write.
Lenses: narrow what an agent looks at for the task in hand. A lens can never widen what the grant allows.
Two rules make it safe to ingest private conversations. It fails closed: anything outside your grants is invisible to you, whatever you ask. And a key given to an agent can carry less access than the person it works for, never more. So you can give a sub-agent access to one client's history, even when the relationship manager it works for sees the whole book.
In practice, a relationship manager's call with a client becomes memories scoped to the people on the call and their team. Their agent can recall it next week. An agent in another division never sees it, whatever it asks.
What a database won't do for you
A database isn't a company brain on its own. A few parts are still your decisions:
Which sources to connect. Calls and chats are where decisions get made. Tickets, the CRM and policy documents are where they get recorded. The brain needs both, and needs to know which one wins when they disagree.
Where to start. One team or one workflow, not the whole organisation on day one.
How you tell people. Employees should know what's captured and who can see it.
So, before you start, make sure you know which conversations hold the knowledge your agents are missing, who should be able to see each of them, and which source is the authority when two of them disagree. If you can answer the second question with permissions you trust, you can ingest much more than you probably think.
For the practical side (where to start, who should own it, the tech stack, deployment models and costs), read How to build a company brain: where to start and what it costs.
Get started
If you're planning a company brain, our solutions engineers can help you pick the first use case and the right deployment model: