Query-Conditioned Reuse Boosts Long-Horizon Agent Trajectory Memory

A recent paper on arXiv (2608.12847) presents a novel approach called query-conditioned reuse (QCR) aimed at leveraging previous agent trajectories for long-horizon tasks. The researchers pinpoint post-retrieval reuse as a critical limitation and introduce a framework that maintains fixed parameters for retrieval, target state, model, decoding, and tool budget while adjusting support. QCR serves as a straightforward method for documenting reusable procedures, bindings, conditions for applicability, and requirements for verification. In tests involving 2,391 target instances from WebArena, WorkArena, and AppWorld, QCR attained an average success rate of 62.3%, surpassing Full Trajectory by 10.7 points and utilizing 48.9% fewer tokens. The full paper can be accessed on arXiv.