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

2026-08-10

AI Summary

This work addresses the challenge faced by long-horizon agents in dynamically retrieving critical evidence from lengthy, interleaved historical trajectories, a task hindered by existing external memory systems that lack the flexibility to compose multiple structured memory views according to task demands. To overcome this limitation, the authors propose a structure-level dynamic selection mechanism that constructs five complementary structured views of agent trajectories and adaptively fuses and queries the most relevant memory subsets based on answer-level weak supervision signals, thereby supporting a frozen answer model. Through end-to-end weakly supervised learning, harness optimization, prior-guided search, and a UCB-based scheduling strategy, the approach effectively balances exploration and exploitation. Evaluated on AMA-Bench, it outperforms the strongest baseline by 8.5% while reducing evidence token consumption by 41%.

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. External memory stores such trajectories as structured representations, yet each structure provides a distinct and incomplete view. 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. 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. Motivated by these findings, we formulate structure-level dynamic selection: selecting and fusing a query-adaptive subset from a library of specialized memory structures. 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. To learn under this weak supervision, MESA employs harness optimization with prior-guided search and UCB-guided scheduling to balance exploration and exploitation. On AMA-Bench, MESA outperforms the strongest baseline by 8.5% while using 41% fewer evidence tokens than the all-structure alternative.

Problem

• long-horizon agents
• evidence selection
• multi-structure memory
• query-adaptive retrieval
• memory composition

Innovation

• dynamic evidence selection
• multi-structure memory
• long-horizon reasoning
• weakly supervised learning
• memory composition