Seeing and Reflecting: Multimodal Memory-Enhanced Agent Collaboration for Recommendation

Existing LLM-based recommender agents often struggle with text-centric limitations, failing to effectively incorporate visual evidence from product images. Furthermore, their memory update mechanisms are frequently coarse-grained, leading to semantic noise and preference drift over time. This paper asks: can a dual-track architecture that separates interpretable agent reasoning from fine-grained multimodal matching improve recommendation accuracy and robustness?

The authors propose MMEACR (Multimodal Memory-Enhanced Agent Collaboration for Recommendation), which employs two distinct tracks:

1. Reasoning Track: User and Item Memory Agents maintain persistent, interpretable multimodal memories. These are updated via an attribute-guided reinforcement-and-reflection mechanism, which uses a predefined semantic attribute space to filter out noise and focus on relevant preference signals.
2. Matching Track: A decoupled multimodal embedding memory processes raw interaction narratives and item images. This preserves fine-grained cross-modal signals that might otherwise be lost during the structured updates in the reasoning track.

The two tracks are integrated using weighted Reciprocal Rank Fusion (RRF), allowing the system to leverage both the explicit reasoning of language agents and the dense similarity signals of multimodal embeddings.

Experiments across three real-world domains (CDs, Cell Phones, and Fashion) demonstrate that MMEACR consistently outperforms competitive LLM-based and agent-based baselines. The framework shows significant improvements in visually grounded scenarios, where the ability to interpret product appearance alongside textual metadata is critical. Ablation studies confirm that the attribute-guided memory evolution is essential for reducing preference drift and maintaining stable, accurate user profiles.