Neural user models are powering today's most prevalent tasks including recommendations and predictions. However, a major short-coming of these models is their lacking ability to model the changing dynamics of user behavior. In this talk, I will present our recent work from KDD 2019 to bridge this gap. Our model, named JODIE, is the next-generation neural user model to learn dynamic embedding trajectories of users. These trajectories are learned to forecast future behavior. The details of the work, including the code, is at https://github.com/srijankr/jodie/.