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

1. Problem Statement

Current agent memory systems have advanced from simple interaction histories to structured external stores and retrieval indices. However, a critical bottleneck remains in long-horizon trajectory memory: the gap between retrieving a relevant past trajectory and effectively reusing it for a new task.

While retrieval identifies a potentially useful past experience, it does not specify how an agent should adapt that experience when users, entities, constraints, or environment states have changed. For short facts or episodes, retrieval often suffices as the reuse mechanism. However, for long trajectories encoding complex workflows, a literal replay often fails because it carries source-specific values (e.g., specific user IDs, file paths, dates) that are obsolete in the target context.

The authors identify post-retrieval reuse as a distinct bottleneck. The core challenge is not merely accessing historical data, but transforming a retrieved source trajectory into a safe, useful support object for a new target task without copying stale bindings or burying the current objective under irrelevant historical noise.

2. Methodology

2.1 Evaluation Framework

The paper introduces a unified evaluation framework designed to isolate the value of the reuse mechanism from retrieval quality.

• Frozen Retrieval: A unified bank of 623 verified historical trajectories (from WebArena, WorkArena, and AppWorld) is constructed. A fixed retriever returns the same top-5 candidates for all compared methods.
• Ranker Selection: A lightweight ranker selects exactly one trajectory from the top-5 candidates before any memory condition is applied. This ensures all methods start with the same "selected experience."
• Target Construction: For each source trajectory, up to four target variants are created. These targets preserve the source's workflow but rewrite specific bindings (entities, users, file paths, dates, etc.) to simulate state changes.
• Controlled Variables: The acting model (DeepSeek-V4-Pro), decoding configuration, tool budget, and target environment state are held fixed. The only variable is how the selected historical experience is represented and delivered to the agent.

2.2 Baselines

The study compares four conditions:

1. No Memory: The agent receives no historical context.
2. Generic Summary: A source-only summary of the trajectory, generated offline without knowledge of the target query.
3. Full Trajectory: The raw, ordered sequence of observations, actions, and tool calls from the selected source trajectory.
4. Query-Conditioned Reuse (QCR): A deliberately simple, target-bound support object.

2.3 The QCR Mechanism

QCR is a minimal intervention that transforms the selected trajectory into a compact note containing four specific fields:

1. Workflow Invariant: The action pattern applicable to the target (e.g., "inspect, validate, modify, verify").
2. Bindings to Re-obtain: A list of values that must be discovered in the current target environment (e.g., "User ID," "File Path"), explicitly preventing the agent from copying source values.
3. Applicability Conditions: Criteria defining when the workflow applies or when reuse should be declined.
4. Verification Guardrail: Checks required to validate the new final state.

QCR is designed not as a claim of a universal memory format, but as a diagnostic tool to test the "reuse hypothesis": whether a target-conditioned transformation improves outcomes compared to raw injection or generic summarization.

3. Key Results

The evaluation was conducted across 2,391 target instances. Key findings include:

• Performance Gains: QCR achieved an average 62.3% Success Rate, outperforming Full Trajectory by 10.7 percentage points and No Memory by a significant margin.
• Efficiency: QCR used 48.9% fewer online tokens (9.4k vs. 18.4k for Full Trajectory) and required fewer API calls, demonstrating that success comes from better representation, not increased context volume.
• Selection Quality: A summary reranking step selected a reusable memory for 94.8% of targets, placing end-task success within 1.8 points of an oracle reusable selector. This confirms that retrieval and selection are effective, but the reuse step is where the remaining gains lie.
• Sensitivity to Length: As trajectory length increases, the utility of direct injection (Full Trajectory) degrades sharply (from +18.4 points for short histories to +2.9 for very long ones). QCR retains 60.3% of its short-trajectory utility even for very long histories, whereas Full Trajectory retains only 15.8%.
• Sensitivity to Binding Shifts: When target bindings differ significantly from the source (large rewrite), Full Trajectory utility drops to +2.2 points, while QCR maintains +20.1 points. QCR reduces "stale-binding errors" (copying old values) from 46.9% (Full Trajectory) to 10.9% under large shifts.

4. Contributions

1. Problem Formulation: The paper defines post-retrieval reuse as a distinct bottleneck in long-horizon agent memory, separating the problem of finding experience from adapting it.
2. Evaluation Protocol: A rigorous framework that holds retrieval, selection, model, and environment fixed to isolate the impact of memory representation and reuse strategies.
3. QCR Framework: A minimal, target-conditioned support mechanism that explicitly separates workflow invariants from source-specific bindings, forcing the agent to re-ground values in the current task.
4. Empirical Evidence: Demonstration that raw trajectory injection becomes less useful as traces grow longer or source-target bindings diverge, while target-bound support preserves utility.

5. Significance and Claims

The authors claim that their work shifts the focus of memory system design from storage and retrieval to adaptation and reuse.

• Memory Utility vs. Retrieval Quality: A high retrieval score does not guarantee that an agent can solve a new task. The value of memory is determined by whether the retrieved experience can be safely transformed into an action plan for the current environment.
• The Role of Summaries: A raw trajectory is not simply a "stronger" version of a short summary; it is an intervention that exposes the agent to both useful procedures and obsolete state. Conversely, a short summary is not automatically useful if it omits necessary preconditions.
• Design Implication: Memory systems should preserve rich, complete records for provenance and evidence in storage, but the acting prompt should contain a compact, target-bound account of the reusable procedure, the bindings to recover, and the necessary checks.

The paper concludes that future systems can employ different retrievers or learned reuse policies, but they must be evaluated on whether they improve target success without importing stale source bindings or hiding the cost of the assistance.