Hermez

Where it came from

Netz decides together, on calls and in chat. Days later nobody remembers what was decided or why, the doubt comes back and the discussion restarts from scratch.

Hermez was born to keep what the house decides, along with the reasoning, and hand it back at the exact moment the question reappears.

On the bench

Memory lives in text files and enters the agent's context on every turn, without anyone having to ask. There are two targets: the profile of whoever is talking, and the facts of the project.

The agent itself decides when to write, and can add, replace or delete. For older material, there is search across past sessions.

The scope decision came early: transient tasks, commits and the day's loose ends stay out. Only stable facts go in, the kind that prevent a repeated question.

Where it fits

It fits a team that loses decisions in conversation and wants recall to happen on its own, at the moment of doubt.

It works well in a small team, with democratic decisions and little bureaucracy. It scales the day memory is treated as a source of truth instead of a log of everything.

What we learned

Scope before storage. An agent that keeps everything retrieves noise, and the useful part disappears in it. The practical rule we adopted: if a piece of information will not prevent a re-discussion thirty days from now, it does not go in.

Automatic injection beats active search. The agent sees the context already in the prompt and uses it without being asked. The day memory requires a "remember that", it has already failed.

Replacing beats accumulating. Swapping the old record keeps memory lean, and only ever adding turns it into a bin. Deleting what has aged is part of the design.

Still open

We still have to evaluate semantic search over the memory, which today goes into the context whole. We also lack an automatic expiry rule for aged facts, and a test on a team of more than five people to find where prompt injection saturates.