<p>Session-based recommendation (SBR) faces critical limitations in existing approaches that hinder recommendation accuracy. Current methods suffer from three fundamental deficiencies: insufficient modeling of temporal dependencies, where standard attention mechanisms treat all sequence positions equally and fail to capture the relative importance of temporal proximity; noisy aggregation in graph neural networks (GNNs), where multi-layer GNN architectures cause over-smoothing and feature information loss as node representations converge to indistinguishable states; and computational inefficiency, where quadratic-complexity self-attention mechanisms introduce noise through indiscriminate all-to-all item connections. To address these challenges, we propose the Mamba-Integrated Spatio-Temporal Attention Graph Convolutional Network (MSTA-GNN). MSTA-GNN tackles temporal dependency neglect through a convolution-based temporal attention mechanism that explicitly encodes temporal ordering and assigns higher weights to temporally closer items. To overcome GNN noise aggregation, we design a single-layer Spatio-Temporal Attention Graph Convolutional Layer (STA-GCL) that achieves higher-order neighbor aggregation while avoiding over-smoothing. For computational efficiency, we replace quadratic self-attention with a linear-complexity Mamba mechanism that selectively filters noise while efficiently capturing long-term dependencies. Additionally, we introduce virtual nodes and a global-level sparse attention encoder to further mitigate noise and capture inter-session dependencies. Extensive experiments on two public datasets demonstrate that MSTA-GNN significantly outperforms state-of-the-art methods, achieving improvements of 10.22%-14.89% in precision metrics and 1.1%-7.66% in MRR metrics.</p>