MemRerank: Preference Memory for Personalized Product Reranking

MemRerank introduces a preference memory framework designed to enhance personalized product reranking in LLM-based shopping agents by distilling user purchase histories into concise signals. The framework employs a benchmark for a 1-in-5 selection task, leveraging reinforcement learning for memory extraction and demonstrating significant improvements in accuracy—up to +10.61 absolute points compared to traditional methods. This advancement highlights the importance of explicit preference memory in improving the effectiveness of personalization strategies in e-commerce applications.