Research into construction of a communication prosthesis for the handicapped requires a system that can model a user's past language acts. We describe a system that continuously adapts to a particular user's needs by maintaining an augmented transition network (ATN) and discuss methods for optimizing and generalizing the information in the network.