Dynamic graphs can accurately portray the evolution of complex networks, so dynamic link prediction methods based on dynamic graphs are becoming mainstream. However, existing methods still have some problems. For the diversity of node topology, the traditional graph model can hardly capture the topological feature of nodes with different roles effectively. For the mutation evolution of the network, the traditional temporal model based on fixed-length time window is difficult to capture such mutation evolution law. To address the above problems, this paper proposes a dynamic link prediction model that combine Hierarchical Convolution and Memory Recall Mechanism (HCN-MR). The model designs a dynamic neighborhood perception algorithm to adaptively adjust the neighborhood perception range of node. On this basis, a hierarchical convolution mechanism is proposed to achieve efficient extraction of complex topological feature. To enhance the ability to capture the mutation evolution laws, the model constructs a dynamic memory bank to store key historical information and perform memory recall. Experiments show that the HCN-MR model exhibits better prediction robustness than the baseline method in complex scenarios of sparse dynamic graphs, especially those with long-tailed node distributions.