Abstract: To improve the representation of long-range dependencies along the depth direction and multivariate collaborative features in long-interval well-log reconstruction, and to reduce the computational cost of attention-based methods in ultra-long well-log sequence modeling, a well-log reconstruction method based on a diffusion model enhanced by Mamba features, namely TSMD, is proposed. The proposed method formulates well-log reconstruction as a conditional diffusion generation process, in which observed logging curves are used as conditional constraints to learn the latent distribution characteristics of logging data. A Mamba sequence modeling structure is introduced into the denoising network, and a Temporal Feature Coupling Module and a Spatial Feature Coupling Module are designed to extract long-range dependencies along the depth direction and correlation features among multivariate logging curves, respectively. Experiments are conducted on real well-log datasets from the Dagang Oilfield in China and Kansas in the United States. The results show that TSMD achieves mean squared error and mean absolute error values of 0.008 and 0.062, respectively, on the Kansas dataset, outperforming the compared methods. Stable reconstruction performance is also obtained on the Dagang Oilfield dataset. The experimental results demonstrate that the proposed method can effectively adapt to long-interval and multivariate well-log reconstruction tasks, providing an effective approach for real well-log data completion and reconstruction.