Introduction Electroencephalography (EEG) based emotion recognition is pivotal for advancing mobile health monitoring and real-time affective interaction. However, current methodologies face a critical trade-off between modeling the complex, multi-scale dynamics of brain activity and maintaining the computational efficiency necessary for edge deployment. Existing approaches often rely on fixed temporal scales and neglect hierarchical spatial connectivity, which limits both classification robustness and scalability in practical settings. Methods To address these challenges, we propose the Multi-Scale Spatiotemporal Graph Mamba (MSGM). Specifically, it employs multi-window temporal segmentation to extract relative power spectral density (rPSD) features, mimicking the brain's multi-scale processing to capture both transient emotional fluctuations and sustained mood. Spatially, it constructs bimodal global and local graphs refined by multi-depth Graph Convolutional Networks (GCNs), intuitively modeling hierarchical brain connectivity rather than isolated sensors. These features are synthesized via a token embedding fusion module and processed by a single-layer MSST-Mamba module, which leverages state-space modeling to ensure linear computational complexity, avoiding Transformer latency bottlenecks to facilitate real-time clinical monitoring. Results Assessed on the SEED, THU-EP, and FACED datasets under subject-independent protocols, MSGM outperforms baseline approaches, attaining competitive accuracy and F1 scores (e.g., 83.43% accuracy and 85.03% F1 score on SEED). Leveraging a single MSST-Mamba layer, MSGM demonstrates robust generalization and efficiency, achieving millisecond-level inference (151 ms) on the NVIDIA Jetson Xavier NX edge device, confirming its suitability for real-time applications. Discussion The capability of MSGM to capture complex spatiotemporal dynamics with low computational overhead highlights its suitability for real-time monitoring and interactive interfaces. By integrating neuroanatomical priors into the selective state-space modeling, the framework effectively maintains spatial intelligence and topological consistency throughout the classification process. This approach not only improves recognition accuracy but also ensures neurophysiologically grounded interpretability. Future research will focus on multimodal integration and further optimization of hierarchical spatial modeling to address the challenges of cross-subject variability. To support research reproducibility, the source code of MSGM will be made available at https://github.com/liuguangyunjizero/MSGM.