Molecule retrieval is critical in medical and scientific research, advancing progress in disease diagnostics and drug development. However, traditional chemical information retrieval tools based on natural language descriptions are limited by the dual challenges of massive molecular data and the lack of detailed textual descriptions for these molecules. To overcome these limitations, cross-modal molecule retrieval task has emerged, enabling the retrieval of target molecules from an unlabeled molecular pool using textual queries. Despite current efforts in feature extraction and aligned semantic space construction, challenges persist, including the complexity of molecular spatial structures, the gap between text and molecular features, and insufficient granularity in cross-modal alignment. To address these issues, this paper proposes a comprehensive cross-modal molecule retrieval framework that optimizes feature extraction and alignment, OFEA-CMR. This framework enhances methods via a triple collaborative mechanism: innovatively using GTN-cen with integrated centrality encoding as the molecular encoder to enhance the modeling capability of molecular spatial structures; designing a shared feature memory based on an exponential moving average mechanism to effectively bridge the semantic differences between text and molecules; and proposing a multi-granularity joint alignment strategy that synergistically optimizes contrastive learning and text-molecule matching tasks, thereby ensuring global alignment between modalities while enhancing fine-grained discrimination of similar samples, ultimately improving retrieval accuracy. Experimental results demonstrate that OFEA-CMR performs excellently across all evaluation metrics, significantly enhancing the performance of cross-modal molecule retrieval.