AttnCompress cut agent trajectory tokens 21.6% and cost 33.6% while beating compression baselines.

It segments trajectories at perplexity spikes to keep code and log syntax intact, uses proxy attention weights to score how relevant each historical block is to current reasoning, and runs a dynamic rolling window that can recall context it previously dropped. On SWE-Bench-Verified and Multi-SWE-Bench it reached 53.17% pass rate above prior compression state-of-the-art. arXiv 2609.08318 The recall behavior is the design point; static pruning can't un-drop a block that turns out to matter three steps later.