Company brain: what it is and how to build one

A company brain is one shared store of what your company knows, which people and AI agents read before they answer or act. It holds the facts you already track, such as clients and contracts, and the material around them: documents, emails, meeting transcripts. It also holds what people said that no field records, such as who decides on a sale or why the team chose one option over another. Agently calls it “a single, shared source of truth that every AI tool and agent in your business draws from…” (agently.dev, read October 2026). Slite calls it “a persistent context layer” (slite.com, read October 2026). Neither definition says the brain must know which client or deal each fact belongs to, and an agent needs that link to find the fact when the client comes up. Standards, an AI-native CRM, builds the brain on the CRM itself. Your records come first, and the memory sits on each record, readable only by the people who can open that record.

A company brain holds five layers of material: records, documents, emails, meetings and memory. Each layer answers a question an agent asks before acting on a client: who is this, what files do we hold, what did the client write, what did the client say in a meeting, and what did nobody write down?

Records carry the structure. Without them, the four other layers have nothing to attach to, and an agent has to guess which “Martin” a mail is about. If your client data still lives in a spreadsheet, start with our guide to building a customer database from your client spreadsheet.

A company brain differs from a wiki, enterprise search and a chat assistant in what it keeps and how it stays current. A wiki keeps what someone wrote down, and enterprise search points you to documents rather than to the client they concern. A chat assistant keeps your conversation and what its own memory holds about you, not the history of each client.

A wiki still has a place: Standards keeps procedures as skills, which you teach once and which every agent applies when the situation comes up.

AI agents use a company brain as the context they load before each answer, and as the place they read and write while they work. In Standards, an agent starts each conversation by reading three profiles: the workspace's rules and mission, the preferences of the person it works for, and its own instructions. When your message mentions a record with @, or when you write from a record's page, the agent receives a card for that record: its main values, up to five related records, its last change and its memories. The agent then searches records by keyword, finds memories and skills by meaning, opens documents and reads mail threads. It acts with the permissions of the person it works for. Auto mode, the default, lets it edit and create without asking; unless that agent's settings say otherwise, before it deletes anything or sends anything outside the firm, it shows you what the action will change and waits for your approval. Manual mode asks before each change to your data.

Agents outside Standards read the same brain. Claude, Codex, Cursor or any MCP client connects to your workspace and reads, searches and writes your records under your permissions. The developers page shows how to plug your own AI tools into Standards.

You build a company brain from the records outward: the structure first, then the material that attaches to it, then the memory and the agents that read it. Steps one to four are set-up work for you, each person, your team and an admin. Once they are done, connected mail, meetings and agent conversations keep feeding the brain with no one copying them in.

A company brain should remember what people said that no record stores, and refuse everything else. Standards' memory refuses field values, statuses, note text and edit history, because the records hold them and copies would drift. It reads five sources: conversations with an agent, connected mailboxes, meeting transcripts, an agent's proposals, and the routine or agent runs that were retried, that had failed before, or in which a person replied after the opening instructions. From a mailbox, it reads only received messages that are flagged important and written to you rather than sent as a newsletter, and never their attachments. It writes each memory as one line: what was said, the source it came from, who said it, the date, and one of five types. Memory files that line on the record the source is already linked to: the records a meeting is linked to, the records matched by a mail's participants, the records mentioned or opened in a conversation. With no linked record and no profile, Standards writes nothing. Agents never write memory themselves: they propose a line, and memory decides in a later review whether to keep it. Memory refuses most of what it reads.

Each Monday, Standards removes the memories whose end date has passed, the context nobody confirmed by its review date, and the lines whose source was deleted.

In a company brain, each person and each agent sees only what they could already open. Standards ties a record's memory to the record: everyone who can see the record reads its memories, and a private deal's memories stay private to the deal's owner. Your mailbox and your private conversations are yours. Memory copies what it learns from them onto a record's page only when the line is a relation, a context or a decision about a workspace record, and only when you can edit that record. A preference or an instruction stays on your own profile. A sensitive line (health, pay, private life, beliefs) never leaves your private scope. A colleague who cannot open the source sees what kind of source it is, whose it is and its date, such as “mail from Bob”, without a link to the mail. The setting “Share with the team what memory learns from my private sources” is on by default, and each person can turn it off.

Harbor Lane Advisors is a fictional twelve-person wealth advisory firm, used here as an example of a company brain at work. The partners create three objects in Standards: households, the contacts in each household with their email addresses, and contracts. Each adviser connects a Gmail inbox, and the meeting bot joins client reviews. During a review, Mrs. Doyle says her daughter will decide on the sale of the family house. Her email address matches her contact record, so the transcript lands there, and memory adds one line to that record's Memories tab: the relation, the meeting it came from, who said it and the date. The adviser running the review typed nothing: the line came from the transcript. Any adviser who can open the record reads that line and sees which meeting it came from; one who can open that meeting follows the link, and one who can edit the record corrects the line if it is wrong.

Two weeks later, the daughter writes to ask about the sale. An adviser opens Mrs. Doyle's record and asks the assistant to draft a reply. The agent reads the record's card, sees the memory, follows its source to the meeting, and drafts an answer addressed to the person who decides. The draft waits for the adviser's approval before it goes out, as anything an agent sends outside the firm does by default.