Integrating robotics in elderly care has shown promise in improving cognitive, physical, and socio-emotional well-being through serious games. However, maintaining engagement and ensuring personalized experiences remains a challenge. This study explores adaptive robotic interaction that dynamically adjusts game content and conversational elements based on user preferences, performance, and difficulty level. This study presents an adaptive robotic system for elderly care, leveraging real-time personalization in serious games through the EBO platform. The system dynamically adjusts game content, conversational elements, and difficulty levels based on user preferences and cognitive abilities. A key innovation is the integration of the CORTEX cognitive architecture with a Deep State Representation model, enabling both therapist-in-the-loop control and autonomous adaptation. The system was evaluated over two months in three elderly care facilities, involving 32 participants with mild to moderate cognitive impairment. Preliminary results indicate a significant improvement in user engagement (4.6/5 satisfaction), motivation (4.4/5), and perceived ease of interaction (4.5/5). These findings contribute to human-centric AI and Industry 5.0 by demonstrating the feasibility of scalable, personalized assistive technologies in real-world settings