Query-Conditioned Reuse Improves AI Agent's Long-Horizon Task Performance

The challenge of reusing past AI agent trajectories, particularly for long-horizon tasks where conditions might change, has been identified as a significant bottleneck. While retrieval can identify relevant past experiences, it doesn't specify how an agent should adapt that experience when user inputs, entities, constraints, or environment states are different. This paper proposes a distinct "post-retrieval reuse" step and an evaluation framework to address this. The researchers introduce Query-Conditioned Reuse (QCR), a simple yet effective method that provides agents with target-bound notes. These notes record reusable procedures, necessary bindings, applicability conditions, and verification requirements. QCR aims to test the hypothesis that structured, context-aware support for reuse is more effective than simply injecting full past trajectories. Evaluations across 2,391 target instances in complex environments like WebArena, WorkArena, and AppWorld showed QCR achieving an average success rate of 62.3%, a 10.7 percentage point improvement over using full trajectories, while consuming 48.9% fewer online tokens. Analysis revealed that direct trajectory injection loses utility as tasks become longer or require significant adaptation, whereas QCR maintains its effectiveness. This framework successfully separates the quality of retrieval from the crucial problem of converting retrieved experience into actionable, safe support for new tasks.