The context-aware systems allow the adaptation of the environment, which depends on the values
perceived by sensors and on the users' preferences detected from user’s behavior. These systems
need to adapt at the users' preferences change so as to generate dynamic systems that improve
their models while they are running. This work proposes a multi-agent system which allows to
learn and to anticipate users' behavior and to adapt at the change into a dynamic mode, and where
storing information from the previous cases during successive execution is no more compulsory.