Traffic forecasting represents a key problem within the domain of spatio-temporal sequence modeling. The primary challenge lies in effectively modeling complex spatio-temporal dependencies. Existing studies typically employ spatio-temporal graph neural networks (STGNNs) to address this problem. However, these graph-based models often rely on static graphs defined by geographical proximity or pre-constructed traffic networks, neglecting the dynamic nature of node correlations within road networks. This limitation hinders the model’s ability to capture evolving spatial dependencies. To overcome this challenge, we propose the Memory Adaptive Dynamic Graph Convolutional Recurrent Network (MADGCRN). First, we introduce a memory network to enhance node representations and construct a memory-adaptive graph, which captures long-term spatial correlations and trends. Simultaneously, a self-attentive mechanism is employed to extract dynamic features from input sequences, generating temporal dependency graphs that model short-term spatial fluctuations. Furthermore, we leverage multi-level feature adaptive integration to balance potential long-term and short-term spatial dependencies in traffic data. Across multiple real-world datasets, our method demonstrates clear advantages over the latest baseline models in performance.