<p>Traffic forecasting plays a crucial role in modern intelligent transportation and urban planning. However, existing methods struggle to effectively address the spatial heterogeneity and temporal non-stationarity inherent in traffic data. Although recent spatio-temporal graph neural networks have shown promising capabilities in capturing spatial and temporal dependencies, most fail to accommodate the hierarchical and dynamic nature of traffic patterns, thereby limiting their adaptability to multi-scale spatio-temporal variations. To address these challenges, we propose a multi-graph memory convolutional recurrent network that innovatively integrates dynamic graph learning, hierarchical memory networks, and graph convolutional recurrent units to model complex and diverse spatio-temporal correlations. Specifically, we introduce a multi-memory network module, where multiple memory networks individually capture spatio-temporal patterns at varying scales, and dynamically generate adaptive graph structures to enhance model generalization under unknown and evolving conditions. Extensive experiments on four benchmark traffic datasets demonstrate that our proposed method consistently outperforms various representative methods, validating its effectiveness in modeling spatio-temporal heterogeneity and non-stationarity.</p>