- In July 2026 Yempik’s context lived in about 137 text files read by Claude: useful, but every change went in without anyone approving it.
- Today an internal platform separates proposals from approved knowledge: email, AI conversations and project code can only propose, and a person decides.
- Every entry says whether it is a decision, a hypothesis or an idea under discussion, so an AI assistant does not mistake an idea for a rule.
- Claude, ChatGPT and Codex connect to the same source with defined permissions: they read and propose, but they do not approve and do not send email.
A folder of files, read by an AI assistant
In the first version our operating context lived in about 137 text files: decisions with a date and a reason, project status, rules, open questions. We published the structure as an open-source project, cowork-os[1], and it still works as a starting point for anyone who wants to get organised. When we opened a session with Claude, the context was already written down.
As we grew, three limits appeared. Anyone, person or assistant, could edit a file and the change immediately became “the right version”. A decision taken looked the same as an idea noted in passing. And each tool read the files its own way: what Claude knew on our computer, ChatGPT on the phone did not.
The problem was never writing things down. It was knowing which ones had been decided, and by whom.
Proposals on one side, approved knowledge on the other
Today we work on an internal platform we built ourselves. It collects what comes out of the work (AI conversations, email, project code, meeting notes) as proposals. A person reads them and decides what goes in. AI assistants connect through MCP (the protocol that connects AI agents to other tools and data)[2], with a single address and access authorised by the person using them.
- 01 · Sources
Where it starts
- Conversations with AI
- Project code
- Notes and meetings
- 02 · Proposals
Waiting
- New decision
- Change to a rule
- Fact about a client
- 03 · Approved
A person decides
- Decisions
- Rules and processes
- Clients and projects
- Open questions
- 04 · Use
Who consults it
- Claude
- ChatGPT
- Codex
- The people on the team
Every entry has a type (decision, rule, process, fact, client, person, glossary term) and is linked to the people, companies and projects it is about. If a new entry contradicts an existing decision, the link stays visible: whoever consults the platform sees the current position and what challenges it.
What the platform enforces, AI assistants included
The rules do not depend on the goodwill of whoever writes: they are built into the software and apply in the same way to a person, an automated procedure and an AI assistant.
- 01
Whoever proposes does not approve
Email, conversations, code and AI assistants can only create proposals. They become approved knowledge when a person accepts them. Every version is kept, with the source it came from.
- 02
Every entry says how certain it is
Idea under discussion, hypothesis, proposal, ratified decision, verified fact. An assistant that answers must say which of these it is quoting, so a September idea does not become an October rule.
- 03
Assistants get permissions, not blank cheques
Each connection has precise permissions: read, propose, start a procedure that is already defined. On email the most it can do is prepare a draft; sending stays with a person.
Before rewriting it, the assistant asked the platform
The previous version of this page told the story of the 137 files. To update it, the AI assistant that works on our website asked the platform for the current position on the Company Brain. The answer came with the decision of 15 September 2026, the entries that limit it and the ones it replaced: the summer pilot subscription, for example, stays on record as history, not as a current offer.
The first draft of this page said the platform is not for sale: the assistant had read that decision too broadly. The founder corrected it (it is no longer a SaaS subscription, but we build and install it for clients as a project) and the correction entered the platform as a proposal awaiting approval. That is why approval stays with a person. The page does not say how many entries the platform holds: the count changes every day.
People decide, agents consult
We do not upload everything and search through it later. Only what a person has approved goes in, and each entry has an owner and a date. AI assistants use it to work with the right context, but people decide the company’s official knowledge.
If you want to start with simple means, the file structure still works: we described it step by step in how to build a company brain on files. We apply the same rules (proposals kept apart from decisions, explicit permissions, a person who approves) to the AI agents we build for clients.
The Yempik case, in practice
Does Yempik really use a company brain?
Yes. Decisions, rules, processes and client information live in an internal platform we built. People approve what becomes official; Claude, ChatGPT and Codex consult it through MCP before they work.
How is it different from the file-based cowork-os?
cowork-os keeps context in text files that an AI assistant reads: it is simple and still a good starting point. The platform adds what files did not give us: proposals kept apart from approved knowledge, the certainty of each entry, links between entries, and different permissions for each connected assistant.
Can an AI assistant change the company’s knowledge?
It can propose a change, not apply it. The proposal waits until a person accepts it. On email the assistant can at most prepare a draft: sending stays with a person.
Can my company have a company brain like this?
Yes, as a project: we build and install it for the individual company, starting from its processes and sources. It is not a SaaS subscription you switch on yourself. The same principles apply to the AI agents we build.
Sources
- [1]cowork-os, Yempik’s open-source repository. github.com
- [2]Model Context Protocol, official documentation. modelcontextprotocol.io
This page is written by Raffaele Zarrelli and Simone Bova, founders of Yempik, with editing done with Claude. It describes Yempik’s platform as of October 2026; the first version, from July 2026, covered the file-based phase. The figure of about 137 files refers to that phase.
Want a company brain for your business?
We build it as a project, with the same rules: what goes in, who approves, what AI assistants can read and propose. Tell us which process to start from.