We present an algorithm for data delivery to mobile sinks in wireless sensor networks. Our algorithm is based on information potentials, which we extend to account for mobility. We show that for local movement along edges in the communication graph, the information potentials can be adapted using a simple iterative distributed computation. However, for non-local movement, the potential field might change significantly, and iterative computation leads to packet loss and packet delivery delays. We address this problem by introducing the mobility graph, which encodes knowledge about likely mobility patterns within the network. The mobility graph can be extracted from training data and is used to predict future relay nodes for the mobile node. Using the mobility graph, we can precompute and efficiently store additional routing states in the network. This enables the algorithm to maintain uninterrupted data streams. We analyze the benefits of computing and maintaining a mobility graph, and show that the information contained therein can be used to improve routing reliability in experiments involving mobile sinks.