The Challenge of Long-Horizon Agent Memory

Agents operating over long time horizons face a critical bottleneck: the 'needle in a haystack' problem. As agents accumulate vast amounts of interaction history, standard retrieval methods—typically based on simple semantic similarity—often fail to capture the specific evidence required for complex, multi-step tasks. This leads to memory noise, where irrelevant information obscures the context needed for effective decision-making.

The MESA Framework: Multi-Structure Evidence Selection

MESA (Multi-Structure Evidence Selection) addresses this by moving beyond flat retrieval. Instead of treating memory as a monolithic block, MESA organizes and retrieves information using a task-adaptive approach that leverages multiple structural representations of the agent's history. By dynamically selecting the most appropriate structure based on the current task requirements, the agent can isolate high-fidelity evidence from long-term logs.

Improving Agent Reasoning and Efficiency

By implementing task-adaptive selection, MESA ensures that the LLM receives a curated 'evidence set' rather than a generic retrieval output. This reduces the cognitive load on the model, minimizes hallucinations caused by irrelevant context, and significantly improves the success rate in long-horizon planning. The framework demonstrates that the way memory is structured and retrieved is as important as the capacity of the memory itself, providing a scalable path for agents to maintain coherence over extended operational periods.