Sequential recommendation predicts users’ future interests by modeling their historical interaction sequences. However, existing methods often neglect the complex dynamic relationship between users’ long-term stable preferences and short-term fluctuating interests when implementing incremental updates for sequential recommendation. Additionally, these methods largely fail to adequately consider the global popularity differences between items and the data sparsity problem, limiting the diversity of recommendation results. This paper proposes GT-DIM: Global-Temporal Dynamic Interest Modeling for Sequential Recommendation to address these issues. The model fully leverages the learnable dynamic characteristics of Kolmogorov-Arnold Networks (KAN) to precisely capture short-term temporal changes and global state information of items based on users’ long-term interactions and implements efficient incremental updates in matrix form through an autoencoder framework. Experimental results on three real-world datasets demonstrate that GT-DIM significantly improves recommendation accuracy. This research opens new avenues for addressing complex temporal dynamics in sequential recommendation.