Company brain ยท interactive demo
One brain, many askers. Current by default, with every earlier version kept.
A company brain captures how an organisation works and serves it to people and agents alike. This demo follows one pricing decision at a fictional company, Acme, through three problems most attempts get wrong: consolidation, permissions and time. SurrealDB holds the graph and enforces access. SurrealDB Agent Memory, beside it, keeps every version and the reason it changed.
Six tools, one company
For decision makers
The model is no longer the bottleneck. The knowledge layer is.
An agent quotes a customer last quarter’s pricing. A new hire spends two weeks reconstructing a decision that was made and then reversed a year earlier. An account manager leaves, and the relationship context leaves with them. None of these is a model failure. Each one happens because no single system holds what the company knows in a form that people and agents can both use.
MIT’s diagnosis is a learning gap rather than a model gap: the tools do not learn how the organisation actually works. A person running a bad search loses ten minutes. An agent working from a bad retrieval takes a wrong action, writes it to a system of record and moves on at machine speed, and its answer looks just as confident as a correct one.
Why most attempts fail
Four failure modes that better retrieval cannot fix
Entities blended together
A search for Phoenix returns an engineering project, an HR programme and an office in Arizona. A model given all three writes one confident answer that mixes them. Consolidation keeps them apart.
Shown in: Ingest and consolidatePermissions added afterwards
A check on the parent document does not stop a semantic search from surfacing a sensitive paragraph out of context. The candidate set has to be secured before anything is ranked.
Shown in: Ask as different peopleFreshness that nobody owns
A vector index returns a stale fact with the same confidence as a current one, because nothing in it records when the fact stopped being true.
Shown in: Travel in timeRanking in place of knowledge
Reranking and hybrid search improve the order of results. They cannot tell a model what is current, what is authoritative or what is private. Those properties have to live in the data.
Evidence at scale
Meta reached 63,000 installs in three months
Meta’s analytics organisation has published how it built an AI second brain for its knowledge workers. It went from zero to more than 63,000 installs and 10,000 daily users in three months, across engineering, product, design, legal, finance, communications and sales. It is the strongest public account in this space because it describes what shipped, what broke and what fixing it took.
Analytics at Meta, How we built an AI second brain for 60k knowledge workers
Governed access came first
Agents reach internal systems through authenticated, scoped MCP servers and CLIs. Meta treated that access as a prerequisite rather than a later addition.
Context arrives in layers
Context-window limits led to progressive disclosure: lean context by default, with detail pulled in only when a task needs it.
Workflows became skills
Versioned, markdown-defined skills such as /start-project turn company knowledge into something an agent can run. Y Combinator’s Summer 2026 Request for Startups lists the company brain as a category it wants to fund, on the same premise.
The market agrees
Glean, the closest public comparison, sells a permissions-aware layer over a company’s fragmented knowledge and has passed $300M in annual recurring revenue.
Timing
When a company brain pays off
Industry matters less than how much decision context lives outside the systems of record, and how many agents need to share it.
Strong fit
- More than one AI agent in production, where the agents have to agree with each other.
- Around 50 people or more, with knowledge spread across five or more tools.
- Customer-facing teams whose relationship context lives in people and calls rather than CRM fields.
- Regulated industries, where who knew what, and when, has to be answerable.
Early
A team of 15 is usually well served by lightweight ambient capture. The full architecture on this page is a different investment, sized for a larger problem.
How to start
- Scope one team, one workflow and one clear before and after.
- Have agents write outcomes back from the pilot onwards, so the brain learns from its own use.
- Settle the permission model before scaling ingestion. It is the most expensive part to retrofit.
Beyond the demo
What is still yours to build?
The database is the substrate: one engine for records, graph, vectors, full-text and permissions. Three parts of a company brain sit above it.
Connectors
Getting events out of Slack, calls, Jira and a CRM, with their access lists intact. Budget weeks per source.
Consolidation policy
Which source is authoritative, and how confident a match must be before two mentions become one entity.
Trust
Ambient capture only works if people know what is captured and who can read it.