Episodic Social Memory via Per-Episode LoRA Adapters for Social Robots

Adapting to individual preferences-including interpersonal distance, formality, and role conventions, is essential for social robots. We introduce a parameter-efficient method for episodic social memory that stores interaction-specific norms as LoRA adapters applied per episode to an open-source dialogue model. We encode episode metadata within a manually defined social feature space, train a distinct LoRA adapter per episode using norm-consistent responses, and at inference retrieve the nearest episode by embedding similarity. We evaluate four configurations: (1) base model (no memory), (2) RAG with episodic text, (3) LoRA-only (activating the retrieved adapter), and (4) combined RAG+LoRA. An independent LLM-as-judge rates outputs for formality, tone, proxemics, and role alignment. Preliminary results on synthetic proxemics and hierarchy tasks indicate that both RAG and episodic LoRA influence behavior, and their combination produces more reliable, user-tailored responses than either component alone.