RSMeM: Knowledge-Enhanced Memory Evolution for Remote Sensing Agents with Systematic Evaluation
This paper introduces RSMeM, a knowledge-enhanced memory evolution framework that improves remote sensing agents' robustness and accuracy by integrating hierarchical domain knowledge with iterative failure-aware experience refinement, achieving significant performance gains on the EarthBench benchmark with minimal additional token overhead.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine you are a detective trying to solve a mystery, but instead of a crime scene, you are looking at the Earth from space. This is the world of Remote Sensing, where scientists use satellites to take pictures of our planet to track things like forest fires, melting ice, or crop health. To make sense of these massive amounts of data, researchers have started using AI Agents—smart computer programs that can think, plan, and use tools (like special software) to analyze the images. However, just like a detective who has never studied forensics, these general AI agents often get confused by the specific jargon and complex rules of geoscience. They might pick the wrong tool or misunderstand a clue, leading to a dead end. The big question is: How do we teach these AI detectives to learn from their mistakes and become true experts without needing a human to hold their hand every single time?
This is exactly what the paper RSMeM (Knowledge-Enhanced Memory Evolution for Remote Sensing Agents) tackles. The authors propose a clever new system that acts like a "super-mentor" for these AI agents. Instead of just letting the AI guess and hope for the best, RSMeM gives the agent two powerful superpowers. First, it provides a hierarchical knowledge base, which is like a massive, organized library of geoscience facts. If the agent needs to know about "land surface temperature," it doesn't just search for any random text; it looks up the specific category in the library to understand exactly what that means and how to measure it. Second, and perhaps more importantly, the system has a memory that evolves. When the agent makes a mistake—say, it tries to use a tool that doesn't exist or mixes up two different types of temperature data—the system doesn't just say "try again." It stops, analyzes why the failure happened, and writes a specific "lesson learned" note into its memory.
Think of it like playing a video game where, instead of just dying and restarting, you get a replay that highlights exactly where you fell off the cliff and gives you a tip on how to jump next time. The paper shows that by combining this organized library with these "lesson notes," the AI agents get significantly better at solving complex tasks. In their tests on a benchmark called EarthBench, the agents using RSMeM improved their accuracy by about 6% (specifically 6.07% on the DeepSeek-V3.2 model) while only using a tiny bit more computer power (less than 1% additional tokens). The authors found that stronger AI models benefited the most because they were better at understanding the lessons and applying them.
The paper explicitly argues against two other common approaches. One is relying on general-purpose AI with no special training, which the authors show leads to "brittle" workflows that break easily when things get complicated. The other is using static, pre-written templates created by human experts. While these templates work for simple tasks, the authors suggest they are too rigid and expensive to maintain, and they don't allow the AI to learn from its own unique mistakes. RSMeM suggests that the best path forward is a dynamic system that starts with expert knowledge but keeps growing smarter every time it encounters a new problem or a failure.
In short, the paper demonstrates that by giving AI agents a structured way to learn from their own errors and grounding them in real domain knowledge, we can make them much more reliable at analyzing our planet. The results suggest that this "memory evolution" approach is a highly efficient way to build expert-level remote sensing agents, turning a clumsy beginner into a seasoned pro with very little extra cost.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.