MESA Improves Long-Horizon AI Agent Memory Retrieval

A new research paper presents MESA (Multi-structure Evidence Selection framework for long-horizon Agent), designed to optimize memory retrieval for AI agents operating over extended periods. Long-horizon agents accumulate vast trajectories of reasoning, actions, and observations, making it challenging to efficiently retrieve relevant evidence for a given query from deep within their history. Existing multi-memory systems often either retrieve too much information or too little, leading to inefficiency or incomplete evidence. MESA addresses this by formulating a structure-level dynamic selection mechanism. It builds five complementary structural views of each agent trajectory and learns to select and fuse a query-specific subset from this library of specialized memory structures. This approach is motivated by findings that optimal memory configurations are typically tailored compositions rather than fixed sets or single structures. The framework learns under weak supervision using harness optimization with prior-guided search and UCB-guided scheduling to balance exploration and exploitation. On the AMA-Bench dataset, MESA demonstrated significant improvements, outperforming the strongest baseline by 8.5% while simultaneously reducing evidence token usage by 41% compared to an all-structure alternative. This makes memory retrieval more efficient and effective for complex, long-horizon tasks.