Time series anomaly detection remains challenging due to complex temporal patterns, signal-to-noise variations, and the need to balance local and global context awareness. Despite notable advancements in existing approaches, several challenges remain, including the inability to dynamically adapt to the inherent characteristics of diverse time series and the difficulty in capturing complex, nonlinear temporal patterns. To address these challenges, we present HAMAMBA, a novel framework that combines hierarchical adaptive attention mechanisms with Mamba state space models for robust anomaly detection. Our approach introduces two core innovations: (1) a multi-granular token routing strategy that dynamically allocates computational resources across temporal contexts, adaptively focusing processing power on informative segments while maintaining broader awareness; and (2) an enhanced Mamba integration that efficiently models long-range dependencies while effectively accommodating non-linear temporal dynamics. Experiments on five benchmark datasets demonstrate that HAMAMBA significantly outperforms state-of-the-art methods and yields F1 scores of 92.84%, 95.22%, 97.05%, 96.51%, and 98.33%. HAMAMBA maintains computational efficiency comparable to recent models while delivering superior detection accuracy, making it well-suited for real-world monitoring applications across diverse domains.