A company leaves behind a lot of information. Messages explain why work changed direction. Documents record plans. Project tools track the work that followed. Putting those records within reach of a model is a useful beginning.
But the question a person asks rarely maps neatly to one document. “Why are we doing this?” may require a decision, the evidence available at the time, and a later conversation that changed its meaning.
Remember the decision, including its conditions
A decision is more useful when the system preserves what it depended on. Who made it? What did they know? Which assumptions might no longer hold?
Consider a team that delays a launch because a customer integration is incomplete. If the integration is later finished, retrieving the old decision without its conditions could mislead someone about the current plan. The system needs to represent change, not simply accumulate text.
This is an illustrative example. It points to a design requirement: a company memory should help distinguish an old statement from a current commitment.
Permissions are part of the meaning
The right answer also depends on who is asking and where the answer will go. Information available to one person may not be appropriate for a team channel.
In the company-brain implementation work, one recurring boundary is the relationship between a claim and its sources. A citation can expose a private title or mailbox identity. Checking the answer’s wording alone will not catch every disclosure.
That pushes permission checks into retrieval, attribution, and delivery. The model should work with context the audience is allowed to use, and the system should recheck that boundary when an output is shared.
Store enough evidence to be corrected
A useful answer needs a path back to the information that supports it. Without that path, a person has to choose between accepting the answer and repeating the research.
Evidence makes correction possible. Someone can identify an outdated source, challenge an interpretation, or explain that a decision has changed. A system that can preserve those corrections has a better basis for its next answer.
What I would test
I would start with a small set of recurring decisions and evaluate the complete path. Can the system find the relevant evidence? Does it distinguish current information from older statements? Does it respect the requester and destination audience? Can a person see what would change the answer?
Then I would measure whether it reduces the work of reaching a supported decision. Retrieval quality matters, but the person still needs to decide and act.
This essay develops an argument from the company-brain project records. Examples are labeled where illustrative. The implementation records support specific permission and evidence checks; they do not establish a measured improvement in company decision-making.