The current popularities of many on-line games and virtual communities are largely based on the diverse reactions of other users and social relationships among the users which computer-controlled characters hardly provide. In this paper, we propose an architecture for implementing an agent called Virtual Inhabitant which can provide diverse and social actions just like those of human-controlled characters. To achieve these human-like actions, the Virtual Inhabitant should behave diversely, live long in the virtual environment, make social relationships with others, keep the relationships in his life time, and have self-updating abilities. To fulfill these requirements, we design the architecture to utilize the agent's full rationality, discrepancies, and other human-likes features such as oblivions, and so on as the essential sources of an agent?s diverse and social actions. We roughly divide the architecture into four sub modules and sophisticatedly design each module in order to support these features effectively. The proposed model can be used for implementing Virtual Inhabitant for on-line games, member of virtual community, pedagogical agents in intelligent tutoring systems with merely selecting and combining the sub modules included in the proposed architecture.