Abstract: Dynamic graph link prediction aims to forecast potential future links between nodes based on existing graph structure data. It currently maps the dynamic graph's structural information into low-dimensional vector representations and extracts the latent evolution patterns within the dynamic graph, and then predicts potential links between nodes by decoding these low-dimensional representations. However, existing dynamic graph link prediction methods face the following issues: 1）forgetting long-term historical information when modeling dynamic graphs; 2）failing to effectively model the dependencies between snapshots when sudden changes occur in the dynamic graph. To address these issues, this study introduces graph contrastive learning methods into temporal dependency learning, proposing a dynamic graph link prediction method based on temporal information contrastive learning. This method performs multi-level contrastive learning on each snapshot within the dynamic graph, constructs new historical information from the results of the contrastive learning, and inputs this into a variational decoder to obtain the dynamic graph link prediction results. Experiments on multiple real-world datasets demonstrate that the proposed method outperforms baseline methods on several metrics and effectively captures the dependencies between different snapshots in the dynamic graph.