Abstract Ambient Assisted Living, smart environments and technology used for elder care, increases independent living time and cuts long‐term care costs. An important requirement for these systems is detecting and informing about abnormal behavior in users' routines. Still, routines may change when certain context conditions hold, which does not mean these changes should concern caregivers. However, changes due to context conditions are not often considered in current proposals. In this paper, we formalize the main concepts related to activities in Ambient Assisted Living, and we introduce contextualized behavior patterns, a long‐term behavior model that considers variability due to context. We adopt a semantic similarity that allows to better detect behavior changes and to understand what makes an observation of daily living different from expected patterns. The results of the experience with three public data sets are promising.