Geoscience research requires complex analysis and domain expertise, with remote sensing (RS) observations as a key foundation. However, existing RS agents built on general-purpose LLMs remain largely domain-agnostic, resulting in brittle and error-prone workflows. Moreover, these failures are seldom consolidated into a reusable experience for subsequent analyses. To address this issue, we introduce RSMeM, a knowledge-enhanced memory evolution mechanism that bootstraps RS agents with pre-distilled domain knowledge and iteratively integrates online experience for robust multi-step tool execution. RSMeM is composed of two components: (i) Hierarchical Knowledge Grounding, which performs taxonomy-aware retrieval over a hierarchical domain corpus to guide planning and tool selection; and (ii) Failure-Aware Experience Refinement, which distills failure-annotated tool-use traces into reusable constraints for next-round tool execution. By iteratively employing these two processes, RS agents can evolve to absorb task-level domain knowledge and effectively translate it into instance-level execution experience. Extensive experiments on EarthBench demonstrate that RSMeM consistently improves tool-use performance and end-to-end accuracy across a diverse set of LLM backbones. Notably, RSMeM achieves a 6% accuracy improvement on DeepSeek-V3.2 with less than 1% additional experience tokens, demonstrating the knowledge density of our distilled experience. All codes and models will be released to support reproducible research.

Memory-augmented assistants must answer questions from long-horizon conversations in which facts are not only missing, but also revised, corrected, or mutually inconsistent. Existing methods largely optimize retrieval, storage, or updating, and many controllers still treat failure primarily as insufficient evidence. We argue that conversational memory QA also requires an explicit model of inconsistency: when remembered claims disagree, a system should not simply retrieve more or guess, but may need to resolve versions, return a qualified answer, ask for clarification, or abstain. We introduce CoMMA, a transparent read-time controller with explicit conflict state for conversational memory QA. CoMMA extracts provenance-aware claims from retrieved memories, groups potentially incompatible claims into conflict clusters, infers lightweight question intent, and chooses among direct answer, qualified answer, clarification, and abstention. In an evaluation centered on the full LongMemEval-Oracle benchmark, a LongMemEval conflict-focused diagnostic slice, and FactConsolidation-20, CoMMA performs best on the main LongMemEval answer-quality comparison, is the only method with non-zero conflict F1 on the conflict-focused diagnostic slice, and ties the strongest write-time overwrite baseline on FactConsolidation-20. These results suggest that making conflict explicit at read time improves both decision quality and the auditability of memory-aware behavior.