One advanced feature of interactive systems in intelligent environments is adaptation: The interface improves itself based on observed user behavior. Therefore, the description of user behavior from observations that the interactive system makes plays a crucial role in adaptive interfaces. This paper presents an architecture for modeling user behavior from these basic events. Different components, such as an interaction and task model, use a common bus system to exchange information through these events. A user modeling component extracts information from events and performs further derivations from existing knowledge, e.g. by means of Markov chains, thus facilitating a comprehensive description of user behavior. A prototype implementation of this architecture, which is applied to an adaptive multimodal digital TV system, demonstrates the feasibility of this approach.