What belongs in a company brain—and what doesn’t?
A research note on useful context, maintained sources, and the difference between remembering and knowing.
A company brain is a useful metaphor until it starts to imply that every file should be copied into one enormous prompt. The more practical question is: what information does this task need, and why should the team trust it?
What the source says
Anthropic’s context-engineering article treats context as a limited resource that must be selected and maintained. It describes approaches that keep references to information and retrieve relevant material when needed, rather than loading everything upfront.
That is a technical argument about what a model sees while working. It is not evidence that a particular knowledge platform will improve your company’s performance. The organizational recommendations below are our interpretation of that distinction.
What we would put within reach
Begin with the sources required by the first workflow: approved procedures, current customer terms, relevant records, and decisions that have actually been adopted. Keep their meaning and ownership attached.
| For each source | Ask |
|---|---|
| Origin | Where did this information come from? |
| Owner | Who can confirm or correct it? |
| Freshness | When was it checked, and what makes it expire? |
| Access | Who is entitled to read or reuse it? |
| Status | Is it a draft, a suggestion, or an approved rule? |
What we would keep out
Do not make personal conversations, restricted employee records, credentials, or unrelated client material generally available just because the system can ingest them. Start from the job and the permissions—not from the storage capacity.
Keep superseded policies distinguishable from current ones. Keep unverified suggestions distinguishable from decisions. A bigger collection is not automatically a more trustworthy collection.
A simple way to test the design
Pick a real question with a known answer. Then create test variations: an expired source, two conflicting sources, a missing source, and a requester without permission. Define what the system should do in each case before running it.
Check whether the response points to the right material, identifies uncertainty, and respects the access boundary. A fluent answer that hides the conflict should not count as a pass.
This is a suggested evaluation exercise. We are not reporting benchmark results or claiming that these checks alone establish safety.
Take it into the next conversation.
The useful company brain is not the one that remembers everything. It is the one whose answers the right people can trace, question, and correct.
Sources & further reading
Anthropic: Effective context engineering for AI agents Selecting and maintaining useful context rather than putting every available document into a prompt.
Source pages checked 19 October 2026. Our examples and recommendations are editorial interpretation, not findings from those publications.Have a workflow in mind?
Let’s talk about the work