Firm memory should take pull requests

A fund memory is the pass reasons, the meeting notes, the LP history, and the fact that someone already looked at this company. The CRM holds the record. The meetings hold the reason. A useful brain returns both, with the evidence it used.

The version that goes wrong writes that memory in place. A model reads a transcript, updates a field, and the next person who asks gets the new sentence as if the firm had agreed it. There is no diff and no author. A wrong line becomes the truth.

Treat the write like a pull request

The model proposes a change. A new fact, a corrected relationship, a pass reason. The proposal shows the field, the current value, the suggested value, and the source line it came from. A person on the investment team merges it. Until that merge, the previous truth stays.

That is the same discipline a codebase already has. Proposed, reviewed, then merged. An agent that commits straight to main is how a firm loses the ability to say what it actually believes.

What a good answer contains

• Sources, each with a place, a date, and an owner.
• A confidence level, and a date by which someone should re-check the answer.
• Permissions. A restricted fact stays restricted in Attio, in Slack, and in the model.
• Open conflicts. If two notes disagree, both stay visible until a person resolves them.

The test we use

Ask why the firm passed on a company. A good answer cites the note and the date. If two notes disagree, it shows both. If the model wants to store a cleaner reason, that cleaner reason is a proposal, and someone has to merge it.

We published a public fixture of this contract in the lab. The companies in it are fictional. The point is the contract: sources, conflicts, and what a merge would have to accept. You can run the context graph fixture and inspect the evidence on each answer. It is the investigation, shown in public, before anyone writes it into a live CRM.