Why does an agent lose track of instructions given earlier in a conversation?

An agent loses track of earlier instructions because it treats each new turn as a chance to revise its answer rather than accumulate the full picture, and that habit gets worse as turns add up. A 2025 study testing top open- and closed-weight models found a 39% average performance drop across six generation tasks when the same instructions were split across a multi-turn conversation instead of given in one prompt. The paper traces this to four habits: models commit to a full answer too early, lean too hard on an earlier wrong attempt instead of correcting it cleanly, overweight the first and last turns while losing the middle, and pad answers with detail that buries what actually changed. None of this is a context window limit; the same conversation usually fits easily in the model’s context. It’s the model’s handling of an instruction unfolding over turns that degrades. Restating the full requirement in one turn typically recovers most of the gap.