<p>Accurately forecasting the remaining useful life (RUL) of lithium-ion batteries (LIBs), a critical component in electric vehicles, is closely tied to essential factors such as system safety, maintenance costs, and resource utilization efficiency. However, achieving accurate RUL predictions remains challenging due to complex nonlinear degradation from variable operating conditions. This study presents a hybrid predictive framework that combines a dual-stream Mamba (DSMamba) and a dynamic filtering frequency mixing learner (DFM) in a synergistic manner. First, Fast Fourier Transform (FFT) decomposes the capacity sequence into low-frequency trends and high-frequency components, forming a dynamic frequency-domain separation framework. A multi-layer perceptron (MLP) generates adaptive time-varying filter coefficients that amplify salient frequency bands and suppress irrelevant noise. This approach solves the frequency-domain mismatch of traditional fixed filters in dynamic degradation scenarios. Second, we propose the DSMamba architecture to extract cross-cycle dynamic features via coupled dual-input cross-connection blocks and state-space models. These two components work together, with the DFM refining the data by filtering out noise and DSMamba extracting meaningful dynamic features from the filtered data. This design also expands the receptive field of battery data representations and significantly enhances long-sequence feature capture capability. On two public datasets, CALCE and NASA, the proposed model achieves the best MAE and RMSE values of 0.0103 and 0.0181. Compared with the MambaSimple model that does not integrate dynamic filtering and dual-stream structure optimization, the proposed optimization model in this paper reduces the error rate by 13.2% and the root mean square error rate by 14.2%.</p>