Main idea

A working session that runs long eventually exceeds what the model can hold at once. At that point it doesn't stop and doesn't warn you: it summarizes the older exchanges to make room, and carries on. The operation is silent and it is performed by the party that is going to use the result — it is the model that decides what deserved to survive.

What gets planed down isn't distributed at random. Whatever was written into a file is indifferent to the operation, since it doesn't live in the conversation. What disappears is exactly the part that hadn't landed anywhere yet: a decision mentioned aloud, a nuance let slip in passing, a trade-off you meant to record later. In other words, compaction only ever destroys undeposited work — but it destroys it entirely, and leaves no mark.

The countermeasure is therefore not to prevent compaction, which is inevitable, but to precede it. Periodically forcing writing to disk over the course of a long session guarantees that what matters is already out of the conversation by the time the model starts choosing. The operation then becomes what it claims to be: a cleanup of the working memory, with no consequence for the memory.

Why it matters

This changes the nature of the risk you think you run with a long conversation. What is usually feared is forgetting between two sessions, which is visible and gets managed; the real risk is a loss inside a single session, undetectable because the conversation carries on exactly as if nothing had been removed.

It also gives a precise moment at which to write, which is neither the end of the session nor an arbitrary rhythm: before the window saturates, that is, while the decision about what matters still belongs to someone.

Nuances and limits

Nothing indicates from the outside that compaction has taken place, or when. The instruction to "push regularly" therefore rests on an estimate, and a dense session can saturate well before the moment you had planned to do it.

And preventive writing only protects what you judged worth writing at the time: the sorting has simply moved from the model to the human, it hasn't disappeared — with the difference that one of the two leaves a trace of its choice.

Open questions

• What observable signal, on the user's side, would tell you that a compaction has just happened and which part of the exchange it covered?