Main idea

What you want to retain is usually filed by theme: one file per subject, one folder per domain. An agent's memory doesn't split that way. It splits according to how long each thing stays true.

Three floors, and each has a different writing rule. The immediate is what the current session has at hand; it disappears with the session. The daily log records what happened today — a bug encountered, a fix attempted, a point left open — and loses its interest within a few days. The long memory, in a long_memory.md kept up to date, keeps only what will still hold in six months: an established technical fact, a project rule, a structuring decision.

So the same piece of information isn't written in the same place depending on the moment. "Transcript extraction was failing this morning" belongs to the day's log; "the hook must be called from /tmp" belongs to the long memory. Confusing the two produces either a journal nobody ever rereads, or a long memory diluted in anecdote.

Why it matters

The filing criterion decides what you can reload. A memory organized by subject forces you to reread everything to know where the work stands; a memory organized by lifespan lets you say "give me the long memory and the last two days" and get exactly what you need to resume.

It is also what makes a forgetting policy possible. A floor known to expire can be archived without debate.

Nuances and limits

A piece of information's lifespan is judged at the moment you write it, which is to say at the worst moment: a decision taken this morning looks structuring, and six weeks later it is only an abandoned implementation detail.

And the floors leak: a daily log always contains a durable fact nobody promoted.

Open questions

• By what signal do you recognize that a line in today's log actually belonged in the long memory?