MESA: Task-Adaptive Multi-Structure Evidence Selection for Long-Horizon Agent Memory

MESA：面向长程智能体记忆的任务自适应多结构证据选择

Abstract: Long-horizon agents accumulate trajectories spanning hundreds of interleaved reasoning, action, and observation steps, where answering a query may depend on evidence buried far back in the history. 摘要： 长程智能体（Long-horizon agents）会积累跨越数百个交替推理、行动和观察步骤的轨迹，其中回答一个查询可能依赖于深埋在历史记录中的证据。

External memory stores such trajectories as structured representations, yet each structure provides a distinct and incomplete view. 外部记忆将这些轨迹存储为结构化表示，但每种结构都提供了一个独特且不完整的视角。

Existing multi-memory systems either read a fixed set of structures for every query, inflating context and introducing noise, or route each query to a single structure, preventing the composition of complementary evidence. 现有的多记忆系统要么为每个查询读取一组固定的结构，这会膨胀上下文并引入噪声；要么将每个查询路由到单一结构，从而阻碍了互补证据的整合。

A controlled analysis on AMA-Bench shows that the optimal memory configuration is typically neither a single structure nor the full union, but a tailored composition of multiple structural memories that varies with query and task demands. 在 AMA-Bench 上进行的对照分析表明，最优的记忆配置通常既不是单一结构，也不是全集，而是根据查询和任务需求而变化的多种结构化记忆的定制组合。

Motivated by these findings, we formulate structure-level dynamic selection: selecting and fusing a query-adaptive subset from a library of specialized memory structures. 受这些发现的启发，我们提出了结构级动态选择：从专门的记忆结构库中选择并融合一个查询自适应的子集。

We propose MESA (a Multi-structure Evidence Selection framework for long-horizon Agent), which builds five complementary structure views of each trajectory and learns from end-to-end answer-level feedback to select and fuse a query-specific subset for a frozen answer model. 我们提出了 MESA（一种面向长程智能体的多结构证据选择框架），它为每个轨迹构建了五个互补的结构视图，并从端到端的答案级反馈中学习，为冻结的答案模型选择并融合一个特定于查询的子集。

To learn under this weak supervision, MESA employs harness optimization with prior-guided search and UCB-guided scheduling to balance exploration and exploitation. 为了在这种弱监督下进行学习，MESA 采用了带有先验引导搜索和 UCB 引导调度的约束优化（harness optimization），以平衡探索与利用。

On AMA-Bench, MESA outperforms the strongest baseline by 8.5% while using 41% fewer evidence tokens than the all-structure alternative. 在 AMA-Bench 上，MESA 的表现比最强的基线模型高出 8.5%，同时比全结构方案减少了 41% 的证据 Token 使用量。