Solid Earth geoscience, geology, geochemistry, and geophysics, is increasingly served by large language models (LLMs) and multimodal LLMs (MLLMs) that combine text, geological maps, hyperspectral imagery, 3D structural models, spatial databases, and quantitative measurements into operational decision pipelines. Existing studies are fragmented across tasks, datasets, and system designs. This survey provides a structured review of LLM/MLLM research in solid-Earth geoscience along three axes: applications, modalities, and Memory-Reason-Act techniques. We identify five clusters of application (information extraction, knowledge Q A and retrieval, multimodal map and 3D interpretation, spatial predictive modelling, and agentic pipelines), characterise four modalities (text, visual, spatial, and quantitative signal) and their encoding gaps, and analyse Memory, Reasoning, and Action design choices with their failure modes. We further examine data resources and bias, and identify the absence of closed-loop drill-outcome validation as the central methodological gap blocking next-generation solid-Earth foundation models.