The availability of enormous event data with timestamps has aroused the study of Dynamic Knowledge Graphs (KGs). In dynamic KGs, temporal link prediction is an important task, which predicts future interactions between entities. Compared with conventional statistic link prediction tasks, temporal link prediction has three main challenges: i) How to deal with new entities that we have not observed before. ii) How to model the temporal evolutionary patterns. iii) How to adapt to the changes in KGs without re-training the model. To deal with these challenges, we present a novel reinforcement learning approach with an update mechanism to integrate temporal information. To predict future events, we train a time-aware agent to navigate the graph conditioned on the input query to find predictive paths. The experimental results indicate a clear improvement over the state-of-the-art methods.