Non-invasive and contactless physiological parameter measurement is a vital research direction in health monitoring, and remote photoplethysmography (rPPG) has become a mainstream technique in this field. Existing rPPG algorithms struggle to balance detection accuracy and lightweight deployment, and suffer from low efficiency in modeling long-range temporal dependencies. To solve these problems, this paper proposes an rPPG-Mamba model. The model adopts depthwise separable convolutions to reduce parameters and computational overhead, and leverages bidirectional Mamba to perform efficient long-range temporal modeling. Temporal difference convolution and 3D channel attention are further integrated to improve feature extraction capability in complex scenarios, enabling accurate rPPG signal extraction from facial videos. Extensive experiments were conducted on three public datasets: PURE, UBFC-rPPG and MMPD. In intra-dataset evaluations, the proposed model achieved a mean absolute error (MAE) of 0.18 and a root mean square error (RMSE) of 0.31 on the PURE dataset, as well as an MAE of 0.43 and an RMSE of 0.68 on the UBFC-rPPG dataset, outperforming all comparison methods. Cross-dataset tests also verified its satisfactory generalization ability. On the MMPD dataset with severe illumination and motion interference, its performance was only slightly worse than that of PhysMamba. Compared with EfficientPhys, the proposed model reduced the parameter count by approximately 89.9%, while its computational complexity was 61.5% lower than that of DeepPhys. The experimental results demonstrate that rPPG-Mamba achieves an excellent trade-off between detection accuracy and lightweight design, and possesses great application potential for non-contact remote physiological measurement.