Policies people will defend
Refunds, privacy, acceptable use, vendor rules. If you would not say it on a recorded call, it does not belong in the company brain.
Company brain
A company brain is not another wiki. It is the approved memory of how you work — the pages a person will stand behind, and the only pages ChatGPT, Claude, Cursor, and your own agents are allowed to use.
What it is
Every company already has knowledge. It lives in Notion, in a shared drive, in Slack threads, in the head of the person who has been here the longest. That is not a company brain. That is a pile. Agents cannot tell which paragraph is current, which is a draft, and which was true last quarter.
A company brain is smaller on purpose. It holds the policies, product facts, processes, and brand rules you would repeat on a recorded call. Each page says when an agent should use it. A person publishes the version other agents may read. Everything else stays a draft.
Memrio is that brain for teams whose AIs already answer customers, write code, and draft copy. One workspace. One MCP URL. A key per agent so support does not see the legal tree, and the coding agent does not quote an unpublished price.
What belongs
If the page is a brainstorm, a meeting note, or a maybe, leave it out. Agents are literal. The company brain should be too.
Refunds, privacy, acceptable use, vendor rules. If you would not say it on a recorded call, it does not belong in the company brain.
What ships, what it costs, what it does not do. Facts the agent should quote, not paraphrase into a launch that never happened.
Onboarding, escalation, offboarding, how a credit is issued. Nested subpages hold the branches. The parent page tells the agent when to open them.
Every page says when to use it. Browse is a catalog of those lines. That is how a company brain stays small enough to trust.
Not a dump
People can wander a wiki. An agent should not. Pointing a model at your entire Drive teaches it to sound confident about pages nobody has reviewed this year.
Wiki or Drive
Company brain in Memrio
How Memrio is the brain
You write pages the way you write for a careful colleague: short, owned, with a note about when they apply. You submit a version. An owner or admin publishes it — or you publish yourself if the workspace allows it. Until then, production agents do not see the change.
Each agent gets a page set and its own API key. The key is the scope. There is no workspace header to forget. Support reads policy. Engineering reads runbooks. A test agent can read drafts. A Writer can edit any page, publish, and create narrower agents later. You can pause any of them.
Agents do not receive embeddings. They call browse, pick pages by title and use-when text, then get_page for the packed contents — instructions, body, attached file text. If they may write, they read the source first, send the whole page back, or nest a subpage under a parent. Every call lands in the request log with the pages it hit.
That loop is the company brain: a person decides what is true, an agent uses only that, and when the agent learns something new it writes a version instead of hiding the fact in a chat.
Who uses a company brain
Support
The support bot and the human on chat both read the same refund, shipping, and account pages. When an exception is approved, it is written back once. The next ticket does not invent a new policy.
Sales
Quote only published SKUs, limits, and competitive language. A new plan lives as a draft until someone publishes it. Cursor and ChatGPT stop promising a feature that is still in review.
Engineering
Incident steps, ownership, and “do not touch” rules sit on pages with use-when instructions. The agent browses, opens the runbook, and follows [CRITICAL] lines instead of guessing from an old README.
Legal & ops
Contracts language, DPA notes, and retention rules are numbered versions. An agent can draft. It cannot silently replace the published page. Review frequency flags anything that has gone stale.
Stand one up
You do not import the company. You publish the pages that already cause damage when they are wrong, then connect the agents that answer those questions today.
01
Name the company brain. The workspace slug becomes the MCP URL every agent will use.
02
Write the policies and facts agents already get wrong. Add use-when instructions. Nest subpages when a topic should split.
03
Issue a key. Point Cursor, Claude, or ChatGPT at the MCP server. The key is the page set — there is no extra header.
04
Turn on edit for a trusted agent, or hand a page to the workspace Writer. Versions wait for a person unless you allow publish.
Questions
Start today
Open a workspace, publish a page, connect an agent. From then on, every tool in the stack reads the same approved memory.