PaperlayerView exact revision on arXiv

Source-linked AI summary

MemRerank: Preference Memory for Personalized Product Reranking

Zhiyuan Peng, Xuyang Wu, Huaixiao Tou, Yi Fang, Yu Gong

arXiv:2603.29247v3cs.CLcs.AIcs.LG

TL;DR

Long purchase histories can overwhelm LLM-based shopping agents and dilute relevant preference signals. MemRerank distills those histories into compact, query-independent memory and trains the extractor with downstream reranking feedback; across retrievers, it reports the strongest macro-average reranking results.

  • Problem

    Naively appending complete purchase histories can strain context windows and dilute useful preference signals, motivating compact preference memory for personalization.

  • Method

    MemRerank extracts compact within- and cross-category preference memory from prior purchases, trains it with GRPO and reranking feedback, and evaluates the full pipeline on fixed top-100 candidate pools.

  • Results

    MemRerank achieves the best macro-average MRR@10 across BM25 and dense retrievers, including a held-out dense retriever, improving the macro average from 7.41 to 8.00.

  • Takeaways & Limitations

    The results suggest that learned preference memory can turn historical shopping behavior into concise, reusable personalization signals for agentic e-commerce search.

  • Takeaways & Limitations

    The evaluation covers textual product metadata and purchase histories from two Amazon-C4 categories, while training and evaluation rely on the proprietary o4-mini API.

Abstract

from arXiv · show

LLM-based shopping agents increasingly rely on long purchase histories and multi-turn interactions for personalization, yet naively appending raw history to prompts is often ineffective due to noise, length, and relevance mismatch. We propose MemRerank, a preference memory framework that distills user purchase history into concise, query-independent signals for personalized product reranking. To study this problem, we build an end-to-end benchmark and evaluation framework centered on an LLM-based \textbf{1-in-5} selection task, which measures both memory quality and downstream reranking utility. We further train the memory extractor with reinforcement learning (RL), using downstream reranking performance as supervision. Experiments with two LLM-based rerankers show that MemRerank consistently outperforms no-memory, raw-history, and off-the-shelf memory baselines, yielding up to \textbf{+10.61} absolute points in 1-in-5 accuracy. These results suggest that explicit preference memory is a practical and effective building block for personalization in agentic e-commerce systems.

1 Introduction

MemRerank addresses the challenge of using long purchase histories for personalized product reranking by distilling them into compact, query-independent preference memory. The paper introduces a benchmark, extractor, and downstream-feedback training setup for evaluating this approach.

  • Naively appending complete purchase histories or transcripts can strain context windows and dilute useful preference signals.
  • MemRerank targets preference memory at the product-reranking stage, where its value can be measured directly against retrieved candidates.
  • The benchmark links Amazon-C4 queries with temporally preceding Amazon-Review-2023 purchases and fixed top-100 candidate pools.
  • The authors release a benchmark containing purchase histories, product-search queries, fixed top-100 candidate pools, and positive-product labels.
  • The preference-memory extractor maps long histories into compact within- and cross-category memories, then uses GRPO with downstream setwise-reranking feedback.
  • Across BM25 and dense retrievers, including a held-out retriever, MemRerank achieves the best macro-average MRR@10 and MRR@5 with shorter memory than raw history and MR.Rec.

2 Related Work

Prior work spans pointwise, pairwise, setwise, and listwise LLM reranking, personalized product-search architectures, and general-purpose agent memory. MemRerank differs by tailoring preference memory specifically to downstream product ranking.

  • Pointwise rerankers score each query-document pair, whereas pairwise and setwise methods compare documents and aggregate wins into rankings.
  • Personalized product-search systems model query, user, and product signals through task-specific architectures such as HEM, ZAM, RTM, CAMI, and PRM.
  • General memory frameworks extract, consolidate, retrieve, or reinforcement-learn compact memories for long-context agents and recommendation reasoning.
  • MemRerank designs preference memory specifically for downstream product ranking rather than general coherence or reasoning.

3 Methodology

MEMRERANK builds a benchmark and two-stage framework that distills purchase histories into reusable preference memory, then trains the extractor with GRPO using downstream five-candidate reranking feedback.

  • 3.1 Dataset Construction: The benchmark links Amazon-C4 queries with temporally preceding Amazon-Review-2023 purchases and evaluates reranking over fixed top-100 candidate pools.Queries are rewritten to preserve shopping intent, and the target positive product is removed from the user’s history before memory construction.
  • 3.2 MEMRERANK Framework: MEMRERANK maps a user’s pre-target purchase history into query-independent within-category and cross-category preference memory.The extracted memory can be reused across future searches without reprocessing the full purchase history for each query.
  • 3.2 MEMRERANK Framework: At inference, a setwise reranker receives the query, extracted memory, and fixed candidate set, comparing five candidates at a time to produce the top-100 ranking.The five-candidate comparison is the atomic operation for both reward collection and final setwise reranking.
  • 3.3 GRPO Training and Reward Design: During GRPO training, sampled memories are evaluated on five-candidate sets containing one positive and four retrieved negatives, with repeated reranker queries providing utility feedback.The reward task uses top-100-aligned retrieval instances and samples negatives from rank buckets spanning ranks 1–100.
  • 3.3 GRPO Training and Reward Design: The GRPO reward combines positive-selection utility with a lightweight deterministic quality term that favors concise, well-formed, non-empty memory and penalizes malformed or unsupported output.Downstream reranking utility remains the main optimization target, while quality checks act as regularization without requiring gold memory annotations.

4 Experimental Setup

The experiments compare MEMRERANK with no-memory, raw-history, direct extraction, and external memory baselines under controlled LLM reranking settings, using MRR@10 as the primary metric.

  • 4.1 Baselines: The study compares direct reranking without memory, raw-history prompting, LLM-extracted memories, external memory systems, and MEMRERANK.External baselines include MR.Rec and Mem0, while the direct extraction comparisons include GPT-5.5 and MEMRERANK without RL.
  • 4.1 Baselines: Table 1 reports top-100 personalized product reranking as MRR@10, with first-stage retrieval metrics separated from o4-mini setwise reranking results.The held-out Linq-Embed-Mistral retriever is excluded from training-negative construction, enabling retriever-generalization evaluation.
  • 4.2 Implementation Details: MEMRERANK uses an RL-trained memory extractor based on Qwen2.5-7B-Instruct, with category-specific memory targets selected using development-set reranking performance.The main table uses within+cross memory for Electronics and within-category memory for Beauty & Personal Care.

5 Experimental Results and Analysis

MEMRERANK delivers stronger category-average reranking than the strongest baselines across the evaluated categories and generalizes to a held-out retriever. Its gains come with compact preference memory, while the best memory target depends on category and quality regularization improves results.

  • Main Results: MEMRERANK improves macro-average MRR@10 from 7.41 to 8.00 across Electronics and Beauty & Personal Care.Category averages rise from 10.59 to 11.25 in Electronics and from 4.33 to 4.75 in Beauty & Personal Care.
  • Retriever Generalization: MEMRERANK achieves the best Linq MRR@10 in both categories, improving Electronics from 13.07 to 14.36 and Beauty & Personal Care from 5.30 to 6.77.Linq-Embed-Mistral is held out from training-negative construction and checkpoint selection.
  • Effectiveness-Memory Tradeoff: MEMRERANK achieves higher MRR@10 than raw history and MR.Rec while using much shorter preference memory.The comparison indicates that improvement does not result simply from increasing prompt length.
  • Ablation Study: The best memory target is category-dependent: Electronics benefits from within+cross memory, whereas Beauty & Personal Care performs better with within-category memory.Electronics reaches 11.25 with M_WC versus 10.71 with M_W; Beauty & Personal Care reaches 4.75 with M_W versus 3.74 with M_WC.
  • Ablation Study: The quality reward improves average MRR@10 in all four category-target comparisons.The gains are +0.73 and +0.89 on Electronics, and +1.40 and +0.51 on Beauty & Personal Care.

6 Conclusion

MEMRERANK provides personalized product reranking through compact, query-independent preference memory rather than direct exposure of long purchase histories. Across two categories and multiple retrievers, it achieves the strongest macro-average MRR@10, including on a held-out dense retriever.

  • Conclusion: MEMRERANK uses compact, query-independent preference memory trained with downstream setwise reranking feedback and memory-quality regularization.The framework trains a Qwen2.5 memory extractor instead of directly feeding long purchase histories to the reranker.
  • Conclusion: MEMRERANK achieves the best macro-average MRR@10 across BM25 and dense retrievers, including a held-out dense retriever not used for training-negative construction.

Limitations

The evaluation is limited to textual metadata and purchase histories from two Amazon-C4 categories, while API-based reranking increases cost at large scale.

  • Limitations: The study focuses on textual product metadata and purchase histories from only two Amazon-C4 categories.Broader evaluation would include more categories, multilingual queries, and multimodal signals such as images.
  • Limitations: Both training reward and final evaluation rely on the proprietary o4-mini API, making reranking more expensive than self-hosted open-weight alternatives at large scale.The paper identifies distilled rerankers and hybrid evaluation protocols as future directions.

A.1 MRR@5

Table 3 reports MRR@5 results for the same test queries used in Table 1.

  • MRR@5: Table 3 reports MRR@5 results for the same test queries as Table 1.

A.2 Prompts

The prompts define memory extraction from users’ purchase histories and LLM-based product selection for personalized recommendation. They include category-specific preference extraction, query rewriting, and reranking with memory.

  • Memory extraction: Purchase items may include category information, product metadata, and optional user review text.The listed metadata includes titles, descriptions, and features.
  • Prompt components: The prompts include separate components for preference-pattern extraction, user-profile extraction, query rewriting, and reranking with memory.These components correspond to the named prompt figures in the appendix.
  • Product selection: The selection prompt asks an intelligent shopping assistant to choose the most relevant product from candidates for a customer search query.This defines the product-selection interaction used by the reranking system.
  • Memory extraction: The memory extractor is instructed to summarize stable preferences for one product category from purchase metadata.The prompt frames the system as a user preference memory extractor for e-commerce recommendation.

A.3 Dataset

The dataset statistics report benchmark-query counts by target category and compare within-category with cross-category purchase-history lengths. Cross-category histories are generally longer.

  • Dataset statistics: Dataset statistics show benchmark-query counts for each target category, with categories having fewer than 100 queries omitted.Categories are sorted by the number of benchmark queries for presentation.
  • Dataset statistics: Cross-category purchase histories are generally longer than within-category histories after excluding the target positive product.The comparison motivates separate modeling of within-category and cross-category histories.
Loading 2603.29247v3…