Long‑horizon language‑model agents continuously accumulate reasoning history during interaction, which inflates context length and inference cost even after earlier decisions have been executed and observed. Unlike static chain‑of‑thought compression, removing past reasoning can alter future actions and the resulting trajectory. We investigate when such reasoning can be safely forgotten.
We introduce Interaction‑Aware Compression for Long‑Horizon Reasoning (ICLR), a training‑free online method that ranks reasoning blocks using a frozen proxy entropy while preserving actions, tool calls, and observations.
On 260 WorkBuddyBench tasks, ICLR raises average reward from 0.699 to 0.718, cuts input tokens by 25.5%, output tokens by 14.4%, and cache‑read tokens by 33.3%.
Ablation studies reveal trajectory amplification: local deletion of reasoning induces nonlinear changes in total computation by reshaping subsequent interaction. Representation probing, activation patching, and controlled trajectory analyses further suggest that once task‑relevant derived state is reliably externalized into code, files, tool outputs, or environmental feedback, historical reasoning becomes more replaceable.
These results characterize agent reasoning as a dynamic working state rather than a permanent interaction history.
Review