Soldiers on mission in areas with unfamiliar cultures must be able to take into account the norms of the local culture when assessing a situation, and must be able to adapt their behavior accordingly. Innovative technologies provide opportunity to train the required skills in an interactive and realistic setting (e.g. serious games, or mixed-reality environments). Such training environments require adequate models that generate the behavior of virtual players. This paper presents an architecture for developing such models. The architecture integrates the Culturally Affected Behavior language (CAB) with modeling behavior as a function of Beliefs, Desires, and Intents (BDI). Culture is defined as norms stored in a separate data file of the agent. During interaction with a human player (e.g. trainee) the agent continuously evaluates whether events are consistent with its cultural norms and responds accordingly. To prevent stereotyped behavior, the agent's behavior was also affected by its personality, defined in terms of Digman's Five Factor Model (1990). The architecture was implemented in JADEX. In a test, agents showed appropriate assessment for their culture and showed behavior consistent with its norms. It is concluded that the architecture may prove important for the development of agent-based training in cultural-aware behavior.