Social isolation affects more than one-fourth of adults aged 65+ in the United States and puts them at a 50% increased risk of dementia and a 32% increased risk of stroke [1]. It is prevalent among vulnerable older adults, especially immigrants and LGBT populations. Older first-generation immigrants are at heightened risk for experiencing loneliness due to language barriers, cultural differences, and racial/ethnic discrimination [2-3]. Social isolation is uniquely amenable to positive outcomes through a range of interventions, including reminiscence therapy [19]. While reminiscence therapy requires a skilled interviewer fluent in the patient’s native language, recent breakthroughs in large language models (LLMs) have demonstrated their ability to engage in human-like, multi-round conversations, unveiling the potential to offer such specialized interventions. In this project, we aim to study how LLMs can help reminiscence therapy toward reducing the expensive human resources. Our project will create an LLM-driven counseling agent (called Memory) aimed at mitigating social isolation by engaging subjects in continuous dialogue in reminiscence therapy. We leverage Retrieval Augmented Generation (RAG) and long-context memory management to meticulously organize memory structures. Our Memory chatbot will be tested by senior immigrants attending local cultural centers. Our hypothesis is that our Memory chatbot can effectively deliver reminiscence therapy for senior immigrants which can be assessed through the improvement in accessibility, user-friendliness, personalization, and engagement. Studies show the effectiveness of using chatbots to deliver reminiscence therapy for early-stage AD patients because their short-term memory is affected but the long-term memory is well preserved [4]. Our long-term goal is to connect isolated seniors based on identified joint interests through our Memory chatbot to build better online communities.