RSMeM: Knowledge-Enhanced Memory Evolution for Remote Sensing Agents
The paper arXiv 2607.24772 presents RSMeM, an innovative mechanism designed to enhance memory evolution for remote sensing (RS) agents. Current RS agents that rely on general-purpose LLMs lack domain specificity, resulting in fragile workflows and frequent errors, with no consolidation of failures into valuable experiences. RSMeM incorporates two key elements: Hierarchical Knowledge Grounding, which utilizes taxonomy-aware retrieval from a structured domain corpus to aid in planning and tool selection, and Failure-Aware Experience Refinement, which converts failure-annotated tool usage into reusable constraints for future tool applications. This mechanism equips RS agents with pre-processed domain knowledge and continuously incorporates real-time experiences to ensure effective multi-step tool execution. The study features a comprehensive evaluation.
Key facts
- RSMeM is a knowledge-enhanced memory evolution mechanism for remote sensing agents.
- It addresses domain-agnostic nature of existing RS agents built on general-purpose LLMs.
- Consists of Hierarchical Knowledge Grounding and Failure-Aware Experience Refinement.
- Hierarchical Knowledge Grounding uses taxonomy-aware retrieval over a hierarchical domain corpus.
- Failure-Aware Experience Refinement distills failure-annotated tool-use traces into reusable constraints.
- Bootstraps RS agents with pre-distilled domain knowledge and iteratively integrates online experience.
- Aims for robust multi-step tool execution.
- Published on arXiv with ID 2607.24772.
Entities
Institutions
- arXiv