Personalized Music Therapy for Dementia Care is a research project that compares three music recommendation pipelines for subjects with dementia: a traditional hybrid recommender, a Vector Retrieval-Augmented Generation (RAG) pipeline, and a Knowledge Graph RAG (KG-RAG) pipeline. All three use the same 7,003 Billboard Hot 100 Year-End songs, the same MMSE-based safety rules, and the same LLM judge, and are evaluated on 260 synthetic personas. On the main set (n = 200), KG-RAG kept the most highly relevant songs and ranked them best (p < 0.001) in both cold and warm sessions.