As the population ages, it is important to maintain the cognitive health of elderlies, which can reduce the cost of care by millions. Through cognitive training can significantly improve or maintain memory, attention, and problem-solving abilities in the elderly. However, traditional cognitive training (non-digital media) methods face issues of low engagement, insufficient motivation, limited accessibility, and scalability. To address these challenges, researchers have developed serious games as a solution. However, current serious games lack personalization and are out of touch with the user environment. Therefore, this paper introduces an adaptive, personalized augmented reality escape game system for cognitive training that addresses these issues. First, we employ a Q&A system to obtains the estimation of the cognitive abilities for the user. Next, we request user to take photo of his current environment. Based on the cognitive abilities and environment, we employ fuzzy system to determine the suitable game programs for the user. Empirical experiments conducted with individuals in their twenties demonstrate that this system can customize cognitive tasks, adapt to the environment, and maintain high user engagement and positive feedback.