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

2026-07-08

AI Summary

This work addresses the limitations of existing large language model (LLM)-based recommender systems, which rely solely on textual inputs and coarse-grained memory updates, thereby failing to leverage visual information and suffering from semantic noise that induces preference drift. To overcome these issues, the authors propose a dual-track memory architecture: the reasoning track employs an attribute-guided reinforcement-reflection mechanism, where user and item memory agents maintain interpretable multimodal memories; the matching track constructs disentangled multimodal embedding memories from interaction texts and product images, and fuses outputs from both tracks via weighted reciprocal rank fusion. This approach jointly enhances reasoning interpretability and preserves cross-modal fine-grained details, significantly outperforming state-of-the-art LLM and agent-based baselines across three real-world scenarios, with particularly notable gains in vision-dependent recommendation tasks.

Abstract

Large language model (LLM)-based agentic recommender systems show promise in modeling user preferences through natural-language reasoning, yet they remain limited by text-centric inputs and coarse-grained memory updates, making agents prone to missing visual evidence, semantic noise, and preference drift. To address these limitations, we propose MMEACR, a Multimodal Memory-Enhanced Agent Collaboration framework for recommendation. MMEACR introduces a dual-track memory architecture that separates interpretable agent reasoning from fine-grained multimodal matching. In the reasoning track, collaborative User and Item Memory Agents maintain persistent multimodal memories and update them through an attribute-guided reinforcement-and-reflection mechanism. In the matching track, a decoupled multi-modal embedding memory is built from raw interaction narratives and item images to preserve detailed cross-modal signals beyond structured memory updates. The two tracks are integrated through weighted Reciprocal Rank Fusion to produce robust and interpretable rankings. Experiments on three real-world domains show that MMEACR achieves strong overall performance against competitive LLM-based and agent-based baselines, with notable gains in visually grounded recommendation scenarios.

Problem

• multimodal recommendation
• LLM-based recommender systems
• memory update
• visual evidence
• preference drift

Innovation

• Multimodal Memory
• Agent Collaboration
• Reinforcement-and-Reflection Mechanism
• Decoupled Embedding Memory
• Reciprocal Rank Fusion