Two partners of a small consultancy review printed documents and a laptop at a bright work table

What is a company brain for AI (and what it really ingests today)

Product8 min read

A company brain is the memory your AI agents query, with citations you can verify. What it is, which sources it ingests today, who it suits and what it is not.

by Equipo MemoryFirst

Table of contents
  1. What a company brain is
  2. What it really ingests today
  3. Citations you can verify
  4. Who it is for and who it is not
  5. What is free and what is Pro
  6. What it is NOT, and where the limits are
  7. How to start without a months-long project
  8. Sources

Last Tuesday someone on your team promised a client a deadline over the phone. On Thursday another person changed the scope in an email thread. On Friday the signed contract says a third thing. When you ask your AI assistant what was agreed with that client, it can only answer from what it can see, and it sees none of this.

That gap between what your company knows and what your AI sees is what a company brain tries to close. We described it as a problem in your company's knowledge is scattered. This article is about the solution: what it is, what it really does today, and what it does not do.

A desk with a phone, a notebook, a printed contract and a coffee cup

What a company brain is

A company brain is a memory layer shared between the people and the AI agents of an organisation. It takes in what happens (events, documents, conversations), splits and indexes it, and returns it when a person or an agent asks, always with the exact reference to where each fact came from.

Three traits set it apart from a search box over documents:

  • It writes as well as reads. Agents can record decisions, blockers and deploys as they happen, not just query what already existed.
  • It understands time. A new fact can supersede an earlier one (May's price against July's), and what was superseded stops showing up in answers.
  • It cites. Every answer points to the source fragment, so you can check it instead of trusting it.

The term took off in 2026. Among the categories Y Combinator asked founders to build in its Summer 2026 Requests for Startups, the "Company Brain" entry (requested by Tom Blomfield) describes an AI-native knowledge system that continuously ingests everything a company produces, so that every employee is as informed as the most knowledgeable person in the organisation. What we build is one specific part of that ambition, not all of it: the memory layer.

What it really ingests today

Our product vision separates what is in production from what is planned. This is the first list, as of 14 November 2026:

SourceStatusDetail
Events through the API (brain/event)In productionDecisions, blockers, deploys, customer contacts
Uploaded filesIn productionPDF, Word, Excel, CSV, images with OCR and WhatsApp exports
Gmail and IMAPIn productionRead-only
Google Calendar and Google DriveIn productionOwn connectors
Dropbox and OneDriveIn productionOwn connectors
SlackIn productionChannels where you invite the bot, no direct messages
TelegramIn productionThrough a bot
Claude Code sessionsIn productionThrough hooks, conversation text only
Call audioPlannedThere is no audio pipeline
Linear, Sentry, WhatsApp BusinessPlannedNo connector yet

Two clarifications matter. First, "voice" exists as a source type, but only for transcripts you bring already written; we do not transcribe audio. Second, Slack syncs incrementally, groups messages per channel and day, and does not propagate later edits or deletions. That is an accepted limit in this version.

Citations you can verify

Every indexed fragment keeps a reference to its origin, and the reference changes with the source type:

TypeWhat the citation holds
DocumentPage and, where relevant, the text box
Voice (transcript)Time range in milliseconds and, if known, the speaker
ChatMessage and thread identifiers
WebAddress and timestamp
OperationsEvent identifier and its origin

So an answer like "the agreed deadline was six weeks" can carry a link to the exact Slack message or the page of the contract. If the citation does not match the answer, you see it. We go deeper in memory you can audit.

A library card catalogue drawer with neatly ordered index cards

How it answers: hybrid retrieval

When you ask, the system combines search by meaning (vectors), search by words (BM25) and relations between entities, and weighs how current each fact is. A language model then writes the answer from the retrieved fragments. If that model is unavailable, the query degrades to the raw passages instead of failing. The mechanism is explained in how retrieval works.

Who it is for and who it is not

It is for:

  • Founders and small teams with many sources and few people to pass information around: law firms, clinics, consultancies, agencies.
  • Teams already working with AI agents (Claude, ChatGPT, Cursor, their own agents) that do not want to build their own memory layer.
  • Anyone who needs to know which source backs an answer, for compliance reasons for example.

It is not for:

  • Anyone looking for a place to write notes: people write elsewhere, the AI reads here.
  • Anyone who needs a self-hosted edition: there is none.
  • Anyone who requires that even inference never leaves the EU: we cannot promise that today (see below).
  • Anyone who wants to transcribe calls: there is no audio pipeline yet.

What is free and what is Pro

According to our pricing page on 14 November 2026:

PlanPriceWhat it includes
Free0 EUR1 project, MCP memory (recall, record, query) and Claude Code hooks, 30-day retention, 1 seat
Pro12 EUR/month billed annually (144 EUR a year)Unlimited projects, unlimited retention, connectors, temporal graph with citations and resolved contradictions, morning brief on Telegram, 1 seat
Team39 EUR/month billed annually (468 EUR a year)Everything in Pro, 5 seats, shared team memory and a librarian queue

VAT included. Paying month to month costs more per month than the annual equivalent; the breakdown is on the pricing page. A practical note: connectors are part of Pro, so on Free the memory is fed by the events your agent records and by the Claude Code hooks.

What it is NOT, and where the limits are

  • It is not a CRM and it does not replace your tools: it plugs into them.
  • It is not a chatbot builder or an agent runtime. It does memory and retrieval, nothing else.
  • It does not train models on your data.
  • Inference is not EU-only. The database, embeddings and original files are in the EU (dedicated infrastructure in Germany, embeddings computed by our own service). But by default the question and the retrieved passages are sent to Cerebras (US) to write the answer. You can bring your own LLM key for queries, although extraction and digests still use the platform chain.
  • Isolation between customers is by query filters plus an optional second barrier at database level. A database per customer is a plan, not a fact.

We would rather say it this way than sell a promise the code does not back.

A developer and a founder talk next to a whiteboard in a small office

How to start without a months-long project

  1. Create an account and a workspace on the sign-up page.
  2. Connect your AI client over MCP (the docs and this article cover it).
  3. If you use Claude Code, follow the guide to memory for Claude Code.
  4. Add one source that really hurts (a client's email, a folder of contracts) and ask something whose answer you already know, to check the citation.

To see how this looks in the day to day of someone with many projects, there is a simulated scenario of a founder with many projects. And if you are comparing options, how to choose a memory layer and the company brain page add context.

Sources

  • Y Combinator, Requests for Startups, "Company Brain" entry (Summer 2026), seen through Modelence's reproduction: https://modelence.com/yc-rfs-summer-2026/company-brain (accessed 2026-11-14). It links to the original: https://www.ycombinator.com/rfs#company-brain
  • On 2026-11-14 the page https://www.ycombinator.com/rfs showed a later batch, so we could not verify the original Summer 2026 text on YC's own site; we only paraphrase what the cited reproduction records (accessed 2026-11-14).
  • Status of sources, citations and data location: internal product documentation checked against the main branch code on 2026-11-13.
  • Prices: MemoryFirst pricing page (accessed 2026-11-14).

Try the brain with one project

The Free plan includes 1 project, MCP memory and the Claude Code hooks. No card needed to start.

Start free

Keep reading