Beyond Retrieval: Query-Conditioned Reuse of Long-Horizon Agent Trajectories

2026-08-13

AI Summary

This work addresses the limitation of existing retrieval mechanisms, which, despite accessing relevant historical trajectories, struggle to effectively reuse them when user goals, environments, or constraints change. To tackle this, the paper introduces the Query-Conditioned Reuse (QCR) framework, which treats trajectory reuse as a critical bottleneck distinct from retrieval. QCR incorporates goal-bound structured memory—encoding reusable procedures, variable bindings, applicability conditions, and validation requirements—to enable safe and efficient experience transfer. Evaluated across 2,391 tasks in WebArena, WorkArena, and AppWorld, QCR achieves a success rate of 62.3%, outperforming full-trajectory baselines by 10.7 percentage points while reducing online token consumption by 48.9%. Moreover, its memory selection matches oracle-level performance on 94.8% of tasks.

Abstract

Retrieval can identify a past trajectory that may matter, yet it does not specify how an acting agent should use that trajectory after users, entities, constraints, or environment state have changed. We identify this post-retrieval reuse step as a distinct bottleneck for long-horizon trajectory memory and formulate an evaluation framework that holds candidate retrieval, target state, model, decoding, and tool budget fixed while varying the support delivered to the agent. We instantiate the framework with query-conditioned reuse (QCR), a deliberately simple target-bound note that records a reusable procedure, bindings to recover, applicability conditions, and verification requirements. QCR serves to test the reuse hypothesis rather than to claim a universally preferred memory format. Across 2,391 target instances in WebArena, WorkArena, and AppWorld, QCR reaches 62.3% average Success, 10.7 points above Full Trajectory, while using 48.9% fewer online tokens. Summary reranking selects a reusable memory for 94.8% of targets, placing end-task Success within 1.8 points of an oracle reusable selector. Analyses by trajectory length and source--target binding shift show that direct trajectory injection loses much of its utility as traces grow longer or source-specific values change, whereas target-bound support preserves a larger share of the measured gain. The resulting framework separates retrieval quality from the problem of turning retrieved experience into safe, useful support for a new task.

Problem

• trajectory reuse
• long-horizon tasks
• memory retrieval
• agent adaptation
• task transfer

Innovation

• Query-Conditioned Reuse
• Trajectory Reuse
• Long-Horizon Agents
• Memory Retrieval
• Task Generalization