Context-aware applications need facilities for recognizing and adapting to context to provide useful and user-centered results. There are several problems to be addressed when building context-aware applications, two of which being how to define and manage all available contextual information and how to distinguish relevant from non-relevant context for a given task. In this paper, we focus on the second problem and propose a context definition and model for context-aware multi-agent system architecture. We exploit this model to build agents that learn to select relevant context and to use it when making decisions.