We propose MarIARel and MarIAToM, two proactive assistants based on Large Language Models (LLMs). They integrate principles from Relevance Theory and Theory of Mind (ToM) to analyze users’ interaction history and initiate personalized, contextually relevant conversations. MarIARel selects users and topics based on relevance, while MarIAToM also considers users’ estimated mental and emotional states. An eight-month experimental study was conducted with 385 patients, who first interacted with a non-proactive assistant that initiated dialogues only when guided by human intervention, and subsequently with MarIARel and MarIAToM. To evaluate the impact of proactive messages, we introduce the metric of response predictability, which quantifies the likelihood of user replies. Results show that MarIAToM, which leverages ToM features, achieved higher response predictability and engagement. We also developed a qualitative assessment method based on emotional impact, applied by human-computer interaction specialists. Experts unanimously found that MarIAToM’s messages were more emotionally engaging and better at sustaining user responses.