ECFMem: A Cognitive-Inspired Dual-Memory Framework for Long-Term Health Dialogue Management in Elderly Care

This paper proposes ECFMem (ElderCare: Fusion Memory Method), a novel memory management framework for large language models (LLMs) in elderly care scenarios. Inspired by the three-stage human memory model (sensory, short-term, and long-term memory), ECFMem dynamically partitions LLM memory into short-term memory-storing recent conversational inputs via attention key-value pairs in a temporary buffer-and long-term memory, which retains critical structured health/habit/relationship data as semantic triplets in an external vector database. To address fragmentation and coherence challenges in elderly dialogues, ECFMem integrates a trainable gating mechanism for contextaware memory fusion and introduces Token-to-Chunk retrieval to optimize short-term recall efficiency. Experiments on elderlyspecific datasets (e.g., EESID) demonstrate that ECFMem significantly enhances memory persistence and response relevance: LLMs augmented with ECFMem achieve up to 85.2% accuracy (a 21% absolute gain over baselines) in long-term health queries and sustain robustness through 1,500+ memory updates. This work bridges cognitive principles with LLM architecture, enabling proactive, personalized elderly companionship.