<p>Traffic flow prediction is a fundamental building block for Intelligent Transportation Systems (ITS), enabling proactive congestion mitigation, adaptive traffic signal control, and reliable route guidance. However, existing spatiotemporal forecasting models still face two practical challenges: they often fail to explicitly reuse cross-period typical traffic patterns (e.g., recurring rush-hour dynamics across days), and their reliance on quadratic-complexity attention limits scalability and real-time deployment in large sensor networks. To address these issues, this study proposes a Hybrid Spatio-Temporal Transformer with Temporal Aggregation and Spatial Memory (HSTT-TASM), aiming to achieve accurate multi-step forecasting while maintaining deployment-friendly efficiency. Specifically, a Temporal Aggregation Convolution (TAC) module captures multi-scale temporal dynamics and reduces redundancy; a Temporal Embedding (TE) module injects learnable periodic and trend priors; and a Temporal Decoupling Layer (TDL) separates long-term trends from short-term event variations through multi-scale decomposition. A dual-channel learner further fuses global temporal attention and local temporal convolution via gating. Moreover, a spatial perception encoder with a memory bank store and retrieves historical spatiotemporal patterns to model evolving inter-node correlations, while kernel-approximated linear-complexity attention significantly reduces computation cost for dense urban networks. Experiments on five real-world traffic datasets demonstrate that HSTT-TASM consistently outperforms state-of-the-art baselines in both forecasting accuracy and computational efficiency, indicating its strong potential as a scalable prediction component for practical ITS decision-support systems.</p>