Brain launches governed company memory for people and AI agents

Brain (heybrain.io) launched a shared, persistent company memory that connects existing knowledge sources, deploys agents on that memory, enforces each source’s permissions on every request, and records access in a tamper-evident, content-blind ledger, with an MCP endpoint so other AI apps can use the same memory.

SOFTWARE desk — companies already want AI on everything they know; what blocks them is proving what the model was allowed to see. Brain packages memory, agents, and a receipt trail in one product.

What the wire says launched. Brain gives people and AI agents a shared, persistent company memory, plus tools for deploying agents and for controlling what each person and each agent can see and do. The wire’s idea is one Brain that gets more useful as the business uses it, then made available to people and to agents without giving everyone the same access. It calls the product something between an AI memory layer, a company knowledge platform, and an agent operating environment. A memory layer, here, is the store of what the company has already said and decided. A knowledge platform is the place that store is searched. An agent operating environment is where software agents are built and run on top of that store. An AI agent, here, is software that can take a series of steps, not only answer one question. These lines are the wire’s. The page does not print the words general availability.

The problem the wire names, as its argument, not as a survey this desk ran. A new ChatGPT conversation starts with limited context. A coding agent may know the repository and not the decisions made in Slack last week. An internal assistant might be connected to company documents and still lack the permissions to tell what one employee should see from what another should see. As companies add more AI, the wire says, they often end up with more fragmented context around it. Fragmented, here, means the knowledge is split across tools that do not share one memory. These sentences are Brain’s.

Where the knowledge comes from, and what the wire says it builds. Brain connects to the places company knowledge already sits, starting with Google Drive and Notion. Slack, GitHub, Box, Telegram, Confluence, Salesforce, and Gmail are listed as in development. In development means those connectors are not the starting pair. Out of the connected sources it builds a knowledge graph of the business: the documents, the people, the projects, and the relationships between them. A knowledge graph, here, is a map of those things and how they connect, not a folder of files. The wire says each source fills in another part of the picture, and each question sharpens what the system understands. It also says that six months in, a team is working with an asset no competitor can buy or copy, because it was assembled from that company’s own history. Six months, and no competitor can copy it, are the wire’s claims. This desk did not wait six months or audit a graph.

How an answer is fetched, and whose cost sentence it is. The wire says loading whole documents into a context window means paying the model to re-read them on every request, so cost climbs as usage grows and accuracy drops once the window fills. A context window is the amount of text the model can look at in one ask. Brain retrieves the passages that answer the question. Answers come from the live source, and the model reads a fraction of the text. A fraction means less than the whole document. The wire does not print a percentage, a token count, or a dollar price for a query. Do not invent one. These lines are Brain’s. This desk did not time a query.

Agents inside the app, as the wire describes them. Brain builds and deploys AI agents inside the app, with the company brain behind them from the first request. The illustrations on the page: a support agent answers from documentation that is current today; sales picks up an account with an agent that has already read the last three calls; overnight, a research agent works through the market folder and reports what changed. Each one runs on the same governed knowledge the humans use. Governed, here, means the rules about who may see what still apply. Every agent carries its own key, its own scope, and a mandate. A key is the credential that says which agent is asking. Scope is what that key is allowed to reach. A mandate is the written limit on what the agent is permitted to do. Actions that carry weight wait for a person to approve them. An agent that needs stopping is stopped with one switch that binds on the next request. The wire says an agent with broad access and no mandate is a script holding production credentials, and that the exposure multiplies with every one a company deploys. It says giving each agent an identity, a limit, and a record is what makes running twenty of them governable. Twenty is that illustration. It is not a customer headcount. This desk did not deploy an agent.

Permissions, and when they are applied. Brain inherits the permissions every source already carries. A private channel stays invisible to anyone outside it. An HR folder stays invisible outside HR. Nothing is copied into a separate store, and nothing is re-shared. IT keeps the access model it already maintains. Those permissions are enforced on every request, before an answer is composed. Where one section of a document is restricted, Brain holds that section back and leaves the rest usable. The wire’s example is that nobody loses a hundred-page document over one paragraph. A hundred pages is the example. It is not a measured document. These lines are the wire’s. This desk did not open a private channel.

The record the wire describes, and the words it uses. Every access writes a record of who asked, which policy applied, and what was withheld, with none of the content stored. Content-blind means the ledger keeps the access and not the text that was read. The about box calls that record a tamper-evident, content-blind ledger. Tamper-evident means a change to the record is supposed to be detectable. The wire says an outside auditor can check that record without taking Brain’s word for it. The about box does not print the word hash-chained. That word is on the product’s MCP page, filed below. Do not treat them as one sentence.

The door for other AI apps, on the wire. AI-native companies get the same memory over an API and an MCP endpoint, so a product team can ship on it in an afternoon. An API is a door other software can call. MCP is the wire’s abbreviation. It does not spell Model Context Protocol on this page. Model Context Protocol is a shared plug so an AI app can call tools and data. An afternoon is the wire’s claim about how fast a team can ship. This desk did not connect an app or ship a product on the endpoint.

Who the wire says was already on it, and the two quotes. Brain went into production at ChainGPT, ChainGPT Pad, and Normies before today’s announcement, where it now serves as the shared knowledge base across every department, from engineering to finance. Before today’s announcement means those three names are early production customers, not a launch-day signup list. Ilan Rakhmanov, identified as CEO at ChainGPT, said that since ChainGPT switched to Brain, the teams closed the knowledge gap, everyone across every department is in sync, and one prompt reaches the information they need, with no more back and forth, document swaps, and meetings. Jayson Burgess, identified as CMO at Brain, said companies have already decided they want AI on everything they know, and what stops them is that nobody can answer what happens when the model retrieves something the person asking was never cleared to see. He called Brain the super app for putting AI to work inside a business: memory, agents, permissions, and proof in one place, so the rollout survives contact with the compliance team. A CMO is the chief marketing officer. Those quotes are on the wire. A quote is not a measured result. The wire also says buyers get a system that was carrying production workloads before it ever carried a price tag. The page does not print a price. The media contact is Jayson Burgess, marketing@aivm.io. The product site the wire names is heybrain.io. The email domain is not a second product name printed in the body.

How a team starts, in the wire’s words. Brain is self-serve and requires no sales process. Sign up at heybrain.io, connect a source, and ask a question. A live demo running on a demo company is available without signup at heybrain.io/demo. Teams evaluating Brain can find comparisons against Microsoft Copilot, Glean, Notion AI, Obsidian, and agent-memory frameworks at heybrain.io/compare. Self-serve and no sales process are the launch words the page prints. The page does not print general availability. This desk did not run the demo and did not score the comparisons. The about box says Brain is a governed, verifiable AI brain for companies, built for teams whose AI rollouts are blocked in security review, including regulated industries, sensitive-data functions such as HR, finance, and legal, and teams deploying autonomous agents. Verifiable, on that box, points at the ledger an auditor can check. Autonomous, here, means an agent that takes steps on its own, inside the mandate.

What the product homepage adds, and where it does not match the wire word for word. heybrain.io, read the same day, describes one shared brain for people and agents. Answers cite the source. Permissions come from the connected source. The FAQ says what is live today: Google Drive and Notion as sources, and Claude, Cursor, and Codex as AI clients over MCP. On the roadmap, not yet available: Slack, Gmail, Box, Telegram, GitHub, Confluence, and Salesforce. That is the same set the wire lists as in development. The order is the homepage’s. The wire’s order is Slack, GitHub, Box, Telegram, Confluence, Salesforce, and Gmail. The FAQ says Brain answers and cites today, and it does not edit documents or create tickets. That limit is the homepage’s. The wire’s agent section does not print it. Do not collapse them. The FAQ’s four controls, which it says are live today: identity, so every request is tied to a named person or a named agent; scope, so an agent’s key limits what it can reach, separately from what its owner can reach; policy, so answers are assembled only from permitted sources and what was withheld is noted; audit, a content-blind record of who asked, what was reached, what was withheld, and when. The wire’s record sentence is who asked, which policy applied, and what was withheld. “What was reached” and “when” are the homepage’s extra words. The FAQ says Brain is hosted by the company, not something a customer runs, and that a free plan has a monthly allowance of credits spent on queries and on indexing. The wire does not print a free plan, a credit, or a price. File the free plan with the homepage.

What the MCP page adds, and only that page. heybrain.io/mcp calls Brain a Model Context Protocol server: one endpoint, governed tools. A key acts strictly as the member who issued it, and the check runs on Brain’s side, not in the client. Every access is appended to a hash-chained record, designed so tampering is detectable, and content-blind by design. Hash-chained means each new line is tied to the line before it, so a quiet edit in the middle would break the chain. That is this page’s word. Tamper-evident is the wire’s about-box word. A key can be rotated or revoked, and the next call stops. Sensitive content is masked before it leaves Brain’s systems. Nothing a customer connects is used to train a model of Brain’s, and a team can bring its own model key so the request runs under that team’s agreement with the model provider. The page says the tools a workspace sees depend on that workspace and its permissions. The tool names stay in the source note. This desk did not call a tool.

Plain English for the rest of the card. Company memory = one shared store of the company’s documents, people, projects, and the links between them, that both people and agents can ask. A knowledge graph = that map. MCP = Model Context Protocol, the shared plug an AI app uses to reach a tool or a data set. The wire prints the abbreviation. The MCP page spells the words. An API = the door other software calls. A key = the credential for one person or one agent. Scope = what that key may reach. A mandate = what the agent is permitted to do. Content-blind = the access record stores who asked and what was withheld, not the text itself. Tamper-evident = the wire’s word for a record a change would show on. Hash-chained = the MCP page’s word for how that record is linked line to line. In development, on the wire, and on the roadmap, on the homepage, = the connectors that are not Google Drive and Notion yet. Self-serve = sign up without a sales process. 14:33 ET = 2:33 p.m. Eastern, the wire’s stamp. ChainGPT, ChainGPT Pad, and Normies = the three early production names. Twenty = the wire’s illustration, not a headcount. This filing is the 23 Sep launch. It is not a login this desk made.

PRIMARY here: GlobeNewswire, 23 Sep 2026, 14:33 ET, datelined Dubai, source Brain — Tier A PRIMARY, the company’s own announcement, carried as sponsored content the wire says it does not endorse. heybrain.io and heybrain.io/mcp are the product pages the wire points at, read the same day, not a second newsroom. The launch description, the Drive and Notion start, the in-development list, the knowledge graph, the passage-retrieval sentence, the agent key, scope, and mandate, the human approval for weighty actions, the one-switch stop, the twenty illustration, the inherited permissions, the withheld-section example, the who-asked record, the tamper-evident content-blind ledger, the API and MCP endpoint, the afternoon ship line, the ChainGPT, ChainGPT Pad, and Normies names, the Rakhmanov and Burgess quotes, the price-tag sentence with no price printed, the self-serve line, the demo URL, the compare list, the about box, and the marketing@aivm.io contact are the wire’s. The live-today Drive, Notion, Claude, Cursor, and Codex lines, the roadmap order, the does-not-edit-documents sentence, the four controls, the what-was-reached and when words, the hosted deployment, and the free plan with credits are the homepage’s. The Model Context Protocol spelling, the key-acts-as-the-member line, the hash-chained record, the revoke-on-the-next-call line, the masking line, and the bring-your-own-model-key line are the MCP page’s. NOT claimed: that the wire prints general availability, a price, a query-cost percentage, a token count, that twenty is a customer count, that hash-chained and tamper-evident are one sentence, that the homepage’s no-edit limit is a sentence on the wire, that this desk signed up, ran the demo, or called a tool, a stock tip, or investment advice. Distinct from the already-filed deepmind-private-ai-memory, wisdomai-live-apps, zerodrift-anchor-3, and form3-agentic-payments.