Long‑horizon language‑model agents continuously accumulate interaction history, which raises computational cost and makes it harder to preserve and reuse relevant information. Existing context‑management approaches mainly address how to compress or retrieve history, but they largely ignore whether the model itself already signals the need for these memory operations before they occur. By examining the hidden state immediately before each agent action, we discover that both compression and recall needs are already encoded in the internal representations. These signals cannot be explained by simple context length or interaction progress and exhibit distinct formation patterns across model depth. Further analysis shows that most memory‑decision information resides in a compact recent context, while selectively restored historical evidence complements the long‑range dependencies missed by the recent context. Building on these insights, we propose Preaction Memory with Evidence Retrieval (PaMER), which combines state‑guided compression with external evidence retrieval. PaMER+ extends this by introducing step‑level evidence selection, retrieving only the historical information required for the current task. Experiments on WorkBuddyBench across multiple context‑management baselines and model backbones demonstrate that our framework substantially reduces context consumption while maintaining competitive task performance.
Review: The paper uncovers latent pre‑action memory signals in LLM agents and offers a practical framework that efficiently balances recent context compression with selective historical retrieval.