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

A recent paper on arXiv (2608.10108) presents MESA, a novel approach for adaptive selection of multi-structure evidence in the memory of long-horizon agents. These agents generate trajectories that involve numerous reasoning, action, and observation steps, often requiring information from distant past events to answer queries. External memory captures these trajectories as structured representations, each offering a unique yet partial perspective. Current multi-memory systems either utilize a static set of structures for all queries, which can lead to excessive context and noise, or direct each query to a single structure, limiting the integration of diverse evidence. Research on AMA-Bench indicates that the best memory setup is usually a customized combination of several structural memories, tailored to specific queries and tasks. In response, the authors propose a dynamic selection method at the structure level, which involves choosing and merging a query-specific subset from a collection of specialized memory structures. This paper highlights a significant challenge in AI memory management, striving to enhance both efficiency and accuracy through the dynamic selection of pertinent memory structures for each query.