Main idea

In a session memory fed by a hook, nobody designates what must survive. At the end of each conversation turn, a lightweight model rereads the last exchange and decides on its own what it records: what goes into the day's log, what moves up into long_memory.md because it judges it durable, what it drops.

The sorting hasn't disappeared for all that — it has changed nature. It was a gesture, it becomes a text: the extraction prompt, the imposed sections, the instruction not to duplicate what already exists. What will be retained in six months is decided today, once, by the phrasing of that criterion, and not exchange by exchange by someone reading.

Hence the shift in effort. Writing a criterion costs more than a one-off gesture, and far less than a thousand gestures. But the criterion also applies to exchanges you didn't anticipate, and nobody notices when it has sorted badly: a delegated memory is silent about what it forgot.

Why it matters

This opens a third way between keeping everything and designating by hand, and it is the only one that scales to daily use: automatic sorting is acceptable only if you can read, correct and version the rule that produces it.

It also says where attention now belongs. Rereading the memory produced isn't only there to correct it: it is the only way to assess the criterion that wrote it.

Nuances and limits

The human doesn't disappear everywhere. In this system, closing a point to follow up on stays a declared act — you send OK-done for a line to move to "Done". What is delegated is the sorting of what comes in; not the verdict on what is finished.

And a written criterion ages like any rule: it was formulated for a project at a given moment, and it keeps applying once the project has changed nature.

Open questions

• How do you measure what an automatic sorting criterion let through, since what is missing shows up nowhere?