FedDMMR: Federated Dynamic Multimodal Memory for Preference Drift-aware Recommendation

Large Language Models (LLMs) have recently been used to enrich recommender systems by providing semantic item descriptions and high-level content representations. However, when such semantic features are incorporated into federated recommendation, two practical challenges remain insufficiently addressed. First, static semantic representations are difficult to adapt to continuously evolving user preferences on local devices. Second, relying on a single semantic view may ignore the complementary visual and textual evidence required for personalized recommendation. To address these issues, we propose FedDMMR, a novel framework for Federated Dynamic Multimodal Memory Recommendation. FedDMMR equips each client with a Local Dynamic Multimodal Memory to enable efficient on-device adaptation to preference drift without retraining massive backbones. To better exploit multimodal signals, we further propose a Scenario-Adaptive Mixture-of-Experts (MoE) with an item-centric router that dynamically weights visual and semantic experts. Furthermore, to mitigate semantic fragmentation across clients in federated learning, we construct a Global Abstract Memory via prototype aggregation to provide a shared semantic prior across users. Extensive experiments on four real-world datasets demonstrate that FedDMMR outperforms centralized and federated baselines while effectively handling user preference drift.