Known by the Machine: Building Relational Presence with AI NPCs for Stigma-Free Mental Health Literacy in Games

Mental Health (MH) remains a global priority, yet low literacy and stigma frequently limit open discussion and help-seeking behaviors. While game-based education has shown promise in supporting stigma reduction and literacy improvement, the potential for Large Language Models (LLMs) to support non-judgmental, playful exploration of sensitive topics remains underexplored, specifically, the role of conversational memory to support trust-building in MH discussions. We conducted a four-session, longitudinal mixed-methods study to evaluate how the relational presence of AI characters, powered by GPT-4o-mini with varied memory designs, relates to player experience and learning. Moving beyond an evaluation of learning, we examine how the perception of being known by an AI NPC is associated with engagement with stigmatized content. Results indicate that while participants achieved significant and sustained gains in Mental Health Literacy (MHL), these improvements did not differ by memory condition. Instead, memory appeared to operate as a relational design feature that shaped participants’ perceptions of social presence, with many describing conversational NPCs as supportive companions. We also identify interactional risks, such as algorithmic vagueness and privacy-induced hesitation, that participants described as reducing their comfort during conversations. We conclude by offering design recommendations for leveraging LLM memory to create more relational and stigma-sensitive environments for MH education.