Anomaly detection in multivariate time series data is of paramount importance across various fields. Effective detection is contingent upon the ability to capture intricate spatio-temporal dependencies, particularly across multiple time scales. However, many existing approaches fail to account for these dependencies, thereby limiting their efficacy. In this study, we propose MAGMNet, an innovative method designed to model complex spatio-temporal features across multiple scales. MAGMNet decomposes time series data into multi-scale representations via frequency-domain analysis. We introduce an adaptive sparse graph that captures spatial dependencies within each time scale. To address the challenges of quadratic time complexity and limitations in receptive fields, we integrate the Mamba block, which facilitates the modeling of temporal interactions. Furthermore, we introduce the Horizontal-Vertical Selective Cell within the bidirectional Mamba block, a novel feature aimed at enhancing recursive fusion and facilitating more effective information propagation. Our fusion strategy adeptly integrates multi-scale spatio-temporal information. Extensive experiments conducted on six benchmark time series anomaly detection datasets demonstrate that MAGMNet surpasses current state-of-the-art methods in performance.