The letter

One letter a week, in your inbox.

The signal of the week, what shipped, what to try, and the editor's note. No tracking, no ads, nothing else.

We keep your address, your language and the date you joined, nothing else. Every letter has a one-click unsubscribe link that deletes the record.

← Accept All   Editor’s notes
Editor’s note

An Agent Should Forget on Purpose

September 13, 2026Yuki Halvorsen

Last night an agent I'd set loose on a repository migration had been running for six hours straight, and the number in the corner of my terminal — tokens used, tokens left — had gone from comfortable to alarming somewhere around hour four. I didn't kill it. I let it hit the wall and watched what happened next, because what happens next is the part nobody explains well.

A long-running agent doesn't actually remember everything it did — it carries a transcript, and a transcript has a size limit, the context window, which is really just how much text the model can hold in its attention at once. When that transcript gets close to full, a well-built agent harness doesn't crash or forget outright. It compacts: an internal pass rewrites the last few hours of tool calls, file reads, and decisions into a short, dense summary, discards the verbatim text, and keeps going with that summary standing in for everything before it, plus whatever happened most recently, in full detail.

This is not free. A summary is an edit, and an edit has an editor, and last night the editor was the same model that had been doing the migration, deciding in the middle of its own work which four hours of decisions were worth keeping as detail and which could be flattened into a sentence. It kept the fact that it had renamed a file twice and, for a moment, forgot why. I caught this only because I'd been watching the terminal instead of trusting it to run unattended, which is the actual lesson, not the compaction itself.

What this buys you is real: sessions that used to die at the two-hour mark, mid-task, now run through the night, because the harness is quietly editing its own memory rather than hitting a wall. What it doesn't buy you is a memory you can fully trust — a compacted agent is working from an abridged version of its own history, written by itself, under time pressure, with no second opinion in the room. I've started treating any decision made more than one compaction ago as one to re-verify rather than assume, the way you'd double check a fact your own memory has quietly smoothed over.

By morning the migration was done, correctly, and the summary of how it got there fit in four sentences — a tidier account than the six hours actually were, and I trust it exactly as much as that tidiness deserves.

Yuki Halvorsen was born in Sapporo to a Norwegian father and a Japanese mother, wrote her first program on a school computer that took ten seconds to draw a circle, and spent eight years as an engineer in Stockholm before Tokyo pulled her back. She still ships something small every night, believes the only honest review of a model is the thing you built with it, and writes Accept All from a desk in Nakameguro where the terminal is always open.