MUSER: Memory-Augmented Multimodal Sequential Recommendation
Abstract
Multimodal Large Language Models offer unprecedented capabilities for sequential recommendation, yet their deployment is fundamentally bottlenecked by the conflict between finite context windows and unbounded interaction histories. Existing methods typically compress multimodal behaviors into textual summaries, inevitably precipitating semantic degradation and interest homogenization. To resolve this, we propose MUSER, a Memory-augmented framework for mUltimodal SEquential Recommendation, built upon a collaborative dual-agent architecture. Specifically, a User Agent maintains a parametric memory via structured input, forget, and rerank operations, guided by a dynamic personality parameter and an Isomorphic Consistency Planning mechanism to ensure long-term preference stability. Concurrently, a Recommendation Agent leverages this stateful memory under a decision-reason bi-objective scheme to jointly optimize recommendation accuracy and high-fidelity textual explanations. Extensive experiments on three real-world benchmarks demonstrate that MUSER significantly outperforms state-of-the-art (SoTA) baselines.