Temporal/dynamic graph link prediction task requires the AI systems to predict the possible future edges based on the observation history of graph events. Existing works for the temporal graph learning mainly use the recurrent neural networks or self-attention mechanism to model the interaction information of graph, and ignore the node attribute features. In particular, they usually adopt a static attribute encoder where the attribute representation can not adjust according the graph structure evolution. To address this issue, we propose Attribute-Empowered Temporal Graph Networks (AE-TGN), which leverages a dynamic attribute leaning module to fine-tune the attribute embedding. In detail, we use the recurrent networks to encode the memory of nodes and the temporal self-attention mechanism to capture the spatial graph information. Next, a trainable multi-layer perceptron is employed to adjust the node attribute representation. Finally, the node attribute embeddings are concatenated with the memory vectors as the temporal embeddings of nodes, and fed into the link prediction decoder. Experimental results on two real-world datasets demonstrates the superiority of our proposal by beating competitive baselines. Specially, AE-TGN presents obvious improvements in the inductive setting, which validates that the attribute encoder can produce accurate node representation for the downstream task.