Due to demographic change, an increase in dementia patients in nursing homes is predicted. The prevention and treatment of dementia patients is time-consuming because the therapy must be individually tailored to the patient’s state of health. As a result of the cost pressure in nursing homes and the shortage of nursing staff, this poses challenges for the nursing area. The use of therapy robots can help to overcome these challenges and strengthen the resilience of elderly people. Previous research has shown that the therapeutic exercises ball throwing, high-five games, and strength exercises are suitable to be supervised and executed by a dementia robot. Essential for the performance and acceptance of such a dementia robot is its interaction capability. Therefore, an adaptive, optimizing, real-time and knowledge based interaction system for the mentioned therapy exercises was researched. This classifies the dementia degree using fuzzy logic and adjusts the exercise parameters with evolutionary algorithms according to the therapy goal. For this purpose, the expert knowledge of caregivers was first collected using a knowledge acquisition and formalized into a knowledge base with a hybrid inference mechanism. A first prototype of the interaction system has already been tested in a real-time robot simulation. After transferring the knowledge-based interaction system to the real dementia robot, it will be evaluated in a test person study.