RSMeM Enhances Remote Sensing AI Agents with Evolving Domain Knowledge.

Researchers have introduced RSMeM, a novel mechanism designed to significantly improve the performance and robustness of remote sensing (RS) AI agents. These agents, often built on general-purpose large language models (LLMs), typically lack the deep domain expertise required for complex geospatial analysis, leading to errors and inefficient workflows. RSMeM addresses this by equipping RS agents with a "memory evolution" system. It first grounds agents with pre-distilled, hierarchical domain knowledge, guiding their planning and tool selection. Crucially, it then iteratively refines this knowledge by distilling insights from past failures into reusable constraints, allowing the agents to learn from experience. This dual approach enables RS agents to absorb task-level domain knowledge and translate it into effective instance-level execution, demonstrating substantial accuracy improvements on benchmarks like EarthBench with minimal additional data.