Abstract: [Objective] Aiming at the challenges that physical signals and geological semantic logic are difficult to synergistically represent in spectral analysis of rocks and minerals, and that existing methods are prone to evidence chain breakage and opaque reasoning in complex scenarios, this paper constructs a geological agent driven by a knowledge graph named RMS KG Agent. It aims to achieve an integrated intelligent analysis covering rock and mineral element description, metallogenic environment inference, and spectral localization of rocks and minerals. [Methods] The system adopts a hierarchically decoupled architecture that integrates a data persistence layer, a computational reasoning layer, and an interactive presentation layer. Centered on Neo4j, it constructs a knowledge graph of five dimensions covering rocks, minerals, chemical compositions, absorption features, and metallogenic environments. Combined with physics aware spectral continuum removal and feature extraction methods such as peak position, peak depth, and full width at half maximum, a task scheduling mechanism is designed. This mechanism prioritizes the knowledge graph, applies rules and physical constraints as secondary steps, and utilizes a large language model for final organization, ultimately developing three core functional modules: soil and rock element description generation, metallogenic environment inference, and spectral localization of rocks and minerals. [Results] The system functionality verification demonstrated that all modules were able to stably produce analytically coherent results with complete structural organization, standardized terminology, and rigorous evidence support. In the spectral localization ablation experiments, the full intelligent-agent configuration achieved Acc@1, Acc@3, and MRR values of 0.8523, 0.8554, and 0.8677, respectively, substantially outperforming the baseline settings that relied solely on peak positions or on the combination of peak positions and absorption depth. In addition, in the comparative experiments on knowledge-driven text generation, the knowledge-graph-enhanced baseline large language model achieved an evidence support ratio of 0.8703, markedly exceeding the 0.4447 obtained by the baseline model, thereby demonstrating a strong balance between factual reliability and rigorous textual expression. [Conclusion] The research indicates that the fusion of multidimensional physical properties and reasoning within a closed loop under knowledge graph constraints are crucial for improving the accuracy of rock and mineral spectral localization, suppressing domain knowledge hallucinations, and enhancing the interpretability of geological reports. The RMS KG Agent effectively bridges the complete technical chain from spectral feature perception, rock and mineral entity matching, and metallogenic environment deduction to professional linguistic expression. This provides a reliable methodological reference for the knowledge based evolution of multimodal geoscience big data and intelligent collaborative exploration between humans and machines.