In this paper, we propose a new and scalable framework that more fully exploits the temporal contextual information of dynamic graph. Dynamic graph node embedding is a technique that embeds nodes in a time-varying graph. When performing representation learning on dynamic graph nodes, it is necessary to preserve not only the topological structure of the graph but also the temporal information. Most existing methods focus more on mining the topological structure of graph and do not fully utilize the temporal information. A convolutional aggregation module that considers temporal context is incorporated, enabling the computation of a node’s embedding to simultaneously consider the information of its neighbors in the current snapshot and the neighboring snapshots. Additionally, the model integrates BiLSTM to aggregate temporal data from both directions. Finally, for each graph snapshot, we introduce a topological reconstruction loss function to generate better embedding for the nodes. We conducted experiments on two tasks and the results of our approach outperform the baseline models.