The company brain

Every account has a brain at /a/<slug>/brain. It is the account's own memory of itself: the refund threshold, what ARR means here, how a client is onboarded, what was decided in June and why, where the signed contracts live, last month's MRR. Agents read it at the start of every run; people read it on Monday morning to see what changed.

The brain is not a document you write. It is derived from what you already have — the documents in the systems you connect, and the conversations you have with Zero and your agents — and confirmed by the people who know. A person always outranks derivation: what you confirm or correct is never overwritten by a later pass, and when a source then changes, the brain asks you to decide rather than deciding for you.

What an entry is

An entry has one of six kinds, and each kind has a shape:

Kind	Shape	Example
Fact	a statement	Refunds above one thousand need the CFO.
Definition	a term, what it means, other names	ARR — annual recurring revenue
Process	a goal and its steps	Onboard a client: sign the MSA, create the project, …
Decision	what was decided, when, by whom, why	Decided 2026-09: move to enterprise — because margins
Location	what, and where it is	Signed contracts are in Drive → Legal → Signed
Metric	a name, a value, a unit, a period, a date	MRR: 12 000 USD (monthly), as of 2026-09-01

Every entry carries at least one source — the document, thread or file it came from, with the passage that supports it — and a status:

• Derived — the machine's claim, with a confidence.
• Confirmed — a person said it is right.
• Corrected — a person changed the text; derivation will not touch it again.
• Pinned — a person wants it in front of every agent, first.
• Retired — gone from every surface and from agents; kept for the record.

Two states a person should act on:

• Needs a decision — two sources say different things, or a source moved under something a person confirmed. The entry leaves the agent block until someone chooses.
• May be out of date — the source changed and the entry has not been re-derived for a month (a week for a metric). Agents see the warning and treat the entry as a lead, not a fact.

Where it comes from

Conversations, always. Account-level memories from your threads become brain entries, visible to the people who were in the conversation. A founder telling Zero we're pivoting to enterprise is knowledge for the founder's own recall until they make it visible to everyone. A thread you set to isolated is never a source.

Connections, when you turn them on. On a connection card (or Brain → Sources) choose Feed the brain. Every night the brain reads the documents that changed since the last pass, derives entries from them, merges what it already knows, and records contradictions as decisions for you. The card shows the cost first: about half a credit per document chunk on the first pass, then only what changes; each account has a daily cap per connection (200 chunks by default, raisable). A cap hit is a clean stop — entries already written stay, the rest wait for tomorrow, and the Sources page says so.

Who sees what

An entry is visible to exactly the people and agents who may open its sources. A document only two members can open produces an entry only those two (and the agents granted the same connection) can see — on the page, in search, through the API and in an agent's prompt. That set is computed when the entry is written and recomputed nightly and whenever a connection's access changes.

The one way an entry's audience grows is an owner or admin choosing Make visible to everyone, which is audited. Correcting an entry does not widen it.

How agents use it

Every agent run starts with a block headed What this company knows: the entries most relevant to the message, pinned first, each with its status and its source, within a fixed budget. Entries that need a decision and retired entries are never in it. An agent that finds nothing is told to say so, not to guess.

Through MCP, brain.search and brain.ask read the brain as the person or agent the token names; brain.ask records an unanswered question as a gap. brain.propose lets an agent suggest an entry — nothing enters the brain until a member confirms the proposal on Brain → Memories. In a conversation, Zero answers what do we know about X from the brain and cites the source.

Gaps

The Gaps page lists the questions nobody could answer — from search, from Zero, from agents, from MCP — most asked first. It is the demand signal for what to connect next. Answer one on the page and a confirmed, account-wide entry is written and the people who asked are told.

The pages

• Brain — the Monday page: a year of activity as a heatmap (a cell per day, the more that changed the darker; open a day to see what it was), what needs a decision, the open gaps, what changed recently, coverage by kind, and which connections feed it.
• Entries — browse and search, filter by kind, status and state. Every entry opens to its shape, its sources with excerpts, its claims if any, and the actions: confirm, correct, pin, retire, make visible to everyone.
• Gaps — the unanswered questions.
• Sources — the connections feeding the brain, their last run (documents seen and changed, entries written, cost, cap hits), the toggle, and the account's daily cap.
• Memories — the agents' and people's own memories (the memory library), one level down.

Cost

Three metered rows, on your account: derive per document chunk extracted, freshness per source checked nightly (a listing, a fraction of a cent), and storage per entry-month. The rates are on the pricing page; the Sources page shows what each connection cost in the last 30 days.