MMEACR framework boosts recommendations with multimodal memory agents

MMEACR addresses the limitations of LLM-based agentic recommenders—text-centric inputs, coarse memory, semantic noise, and preference drift—by introducing a dual-track architecture. The reasoning track uses collaborative User and Item Memory Agents that maintain persistent multimodal memories updated via attribute-guided reinforcement-and-reflection. The matching track builds a decoupled multimodal embedding memory from raw interaction narratives and item images, preserving cross-modal signals beyond structured updates. The two tracks are integrated through weighted Reciprocal Rank Fusion for robust, interpretable rankings.

Experiments on three real-world domains demonstrate MMEACR's strong overall performance against competitive LLM-based and agent-based baselines, with notable gains in visually grounded recommendation scenarios. The framework effectively combines interpretable reasoning with fine-grained multimodal matching, offering a practical approach for building more accurate and robust recommendation systems that leverage both text and images.