Role-playing is an emerging application of large language models (LLMs), allowing users to be immersed in conversations with virtual characters by mimicking their tones and background knowledge. It can be applied in various scenarios such as gaming and virtual reality systems. However, existing methods ignore two challenges: (1) ignoring the relationship with the role played by the user will diminish the immersive experience of the user; (2) insufficient understanding of the character’s background knowledge may lead to inconsistent dialogue. In this paper, we introduce the Duplex Relationship Modeling based Role-play framework (DRMR), a novel role-playing framework designed to enhance the immersion of user when interacting with the role-play model. We first propose a graph-based relationship modeling method, utilizing graph structures to model the duplex relationship between the user and the model’s played characters. In order to better extract useful personalized information about roles from historical dialogues, we construct a role memory consisting of the description of the duplex relationship. To avoid generating an inconsistent response, we iteratively verify the generated response by updating the role memory according to the current dialogue context. Extensive experiments on benchmark dataset demonstrate the effectiveness of DRMR in enhancing user immersion in role-playing interactions.