How context grows when everything is kept versus when a harness decides what to carry forward.

Something has to decide what it forgets

The loop feeds every observation back into the prompt, and the prompt has a hard ceiling. So on a long task the real question is not what the model can read — it is what the harness chooses to carry forward.

steps taken so far8
keep every page it ever read
keep four things only

the four things worth keeping

Dropping the page but keeping the quote is the whole idea. A page it has finished with is dead weight; the sentence it pulled out of that page is the reason it read it at all.