Social robots can assist older adults in their daily life. Verbal conversation is a natural and convenient way for older adults to interact with social robots. However, most of the existing conversation-based robot services, such as medication reminders, are rule-based systems. These systems require many hand-crafted rules and a significant amount of expert knowledge, therefore they cannot adapt to older adults’ characteristics and dialog history. There are many reinforcement learning (RL) based methods for task-oriented dialogues, but they mainly focus on completing the tasks through text-based conversations. Those methods cannot be directly used for elderly care applications involving human–robot interactions (HRI). Considering the above shortcomings, we proposed a dialog system adaptation method (DSAM) for social robots. The DSAM is based on reinforcement learning which considers the characteristics of older adults, the dialog history and user preference to adapt the dialog policy and improve the dialog module. We implemented DSAM in our custom-made ASCCBot social robot. To evaluate DSAM, we firstly tested the dialog agent which was trained by a user simulator with different settings. The results show that the obtained agent achieves a good result with the desired dialog flow compared to the baseline agent. Based on the obtained dialog policy, the adaptation process is evaluated. The results show that with a good success rate, the number of dialog turns is decreased and the NLU module performance is improved by the adaptation process, which proves the effectiveness of DSAM. We also tested DSAM with human subjects. The results show that the average adaptation success rate is 94.7% and the preference distance reaches 0 after 6 rounds of adaptation while creating reminders successfully with a limited amount of user feedback. • A robotic dialog adaptation system for medication reminder in elderly care. • Reinforcement learning based methods for task-oriented dialogues. • Characteristics of older adults, dialog history and user preference considered in adaptation. • User simulator and user modeling combined to adapt the dialog agent.