Originally written in French. Translated by AI — the meaning has been preserved, not the prose.
Main idea
Asking an agent to maintain a summary per topic gives you exactly what you asked for: a record that is always accurate. On a building project, it ends up saying "project complete, handover satisfactory", and that is true.
Nobody deleted what has disappeared: each update overwrote the one before. When did the project start, when was it delayed, on what date did an amendment attach itself to it — the current state carries none of those answers, and was never designed to. An aggregate answers "where do things stand?"; it does not answer "how did we get here?", and the second question always comes.
The two views therefore cannot be derived from one another. A chronological log kept alongside the aggregate — one dated line per event, never rewritten — is not redundancy: it is the subject's other half. One is read in ten seconds to understand, the other is climbed back up to reconstruct. Trying to keep only one amounts to choosing in advance which of the two questions you forbid yourself to ask.
Why it matters
This names a loss that nothing signals. An aggregation system produces no error when it overwrites the history: it produces an accurate record, and that is precisely why the gap takes months to show — it appears only at the first dispute, the first handover of the file, the first newcomer.
It also gives a simple design criterion, before writing a single directive: asking an agent to maintain a state is asking it to destroy at every turn. If the trajectory matters, it must have a writing place of its own.
Nuances and limits
The log carries a reading cost the aggregate doesn't: it grows without limit, and on a topic that runs for years it turns back into the email chain you were trying to escape.
And not every topic deserves both views. A list of contractors has no trajectory — it grows without ever contradicting itself, and attaching a journal to it would produce nothing but volume.
Open questions
- By what signal do you recognize that a topic has stopped having a trajectory, and that its journal can be closed?