Abstract: To address the engineering challenges in long-interval multi-variable logging curve reconstruction, including insufficient representation of long-range dependence in ultra-long sequences, inadequate mining of multi-curve coupling features, and the high computational cost and low operational efficiency of traditional attention mechanism based models, a logging curve reconstruction method that integrated a diffusion model with Mamba feature enhancement (TSMD) was proposed. The proposed method formulated the logging curve reconstruction task as a conditional diffusion generation problem, leveraging fully observed curves as constraints to learn the underlying distribution of real logging data. A lightweight Mamba state space modeling structure was introduced into the reverse denoising network of diffusion. Additionally, a temporal feature coupling module (TFCM) and a spatial feature coupling module (SFCM) were designed to precisely capture the continuous long-range evolution features of formations along the well depth dimension and the nonlinear synergistic correlation features among multi-variable logging curves, respectively. Comparative experiments conducted on two real logging datasets from the Dagang Oilfield in China and Kansas in the United States demonstrated that TSMD achieved a mean squared error of only 0.008 and a mean absolute error of only 0.062 on the Kansas public dataset, significantly outperforming traditional machine learning, deep learning, and mainstream generative compared models. On the Dagang Oilfield measured dataset, stable and high-accuracy reconstruction performance was also obtained. Ablation studies confirmed that the Mamba architecture and the dual feature coupling modules effectively reduced model computational complexity and enhanced spatiotemporal feature representation capability. The research findings provide efficient and reliable intelligent technical support for missing data imputation and accurate reconstruction in long-interval, multi-variable, and large-scale logging data.