Automated depression detection from audio-visual cues faces a key challenge: subtle behavioral indicators are often masked by dominant low-frequency signals. To address this, we propose WavGateMamba, a novel architecture integrating wavelet analysis with state space models. Our method employs Discrete Wavelet Transform (DWT) to decompose inputs into multi-scale frequency components, a gating mechanism for dynamic feature recalibration, and Mamba blocks with shared state-space matrices for cross-modal modeling. Extensive experiments show our approach significantly outperforms existing methods, achieving consistent improvements of 3–4% on key evaluation metrics including Accuracy, Precision, Recall, and F1-Score. This work presents an effective paradigm for incorporating frequency-domain inductive bias into behavioral analysis, enabling better capture of subtle depressive cues.