MemRerank: Preference Memory for Personalized Product Reranking

• Naively appending purchase histories or transcripts can strain context windows and dilute useful preference signals, motivating compact, structured memory of within- and cross-category shopping patterns.Source: §1 · p. 1
• MEMRERANK distills purchase histories into query-independent within-category or within- and cross-category preference memory, reusable across future searches.The memory extractor is post-trained with GRPO (Shao et al., 2024) using feedback from a downstream setwise LLM reranker.Source: §1 · pp. 1–2
• The benchmark pairs Amazon-C4 queries with temporally preceding Amazon-Review-2023 purchase histories and fixed top-100 candidate pools from the full product corpus.It includes positive-product labels for personalized product reranking.Source: §1 · p. 2
• Across BM25 and dense retrievers, including a held-out retriever, MEMRERANK achieves the best macro-average MRR@10 and MRR@5 with much shorter memory than raw history and MR.Rec.Source: §1 · p. 2

2.1 LLM-based Rerankers

• Setwise reranking compares document groups, and this work improves its input with concise preference memory rather than changing the ranker.The related work also distinguishes pointwise scoring and listwise ranking approaches; the authors choose setwise reranking for efficiency (Peng et al., 2025).Source: §2.1 · p. 2

2.2 Personalized Product Search and Reranking

• Prior personalized search and reranking methods learn task-specific models, whereas this work uses query-independent preference memory with a fixed setwise LLM reranker to isolate memory-construction strategies.The comparisons use the same candidate pool.Source: §2.2 · p. 2

2.3 Memory in Agentic Systems

• Agent-memory systems support long-horizon reasoning, recommendation, or dialogue, while this work tailors preference memory specifically to downstream product ranking.Examples include Mem0 (Chhikara et al., 2025), MemAgent (Yu et al., 2025), and recommendation-focused MR.Rec (Huang et al., 2025b).Source: §2.3 · pp. 2–3

3.1 Dataset Construction

• The benchmark pairs rewritten Amazon-C4 queries with each user’s pre-target Amazon-Review-2023 purchases, excludes the positive item, and evaluates fixed top-100 pools from four retrievers.Queries are rewritten to preserve shopping intent while removing review-specific details; products are represented by title and description. Amazon-C4 is cited as Hou et al. (2024), and the rewrite follows MR.Rec (Huang et al., 2025b).Source: §3.1 · p. 3

3.2 MEMRERANK Framework

• The extractor maps purchase history to query-independent memory, either within-category only or split into within- and cross-category preferences, for reuse across future searches.These targets are M_W and M_WC, respectively.Source: §3.2 · pp. 3–4
• At inference, a setwise reranker compares five-candidate subsets from the fixed top-100 pool and aggregates decisions into a full ranking; reported metrics cover the full pool.The same five-candidate comparison is the atomic operation for reward collection and final reranking.Source: §3.2 · p. 3

3.3 GRPO Training and Reward Design

• Training uses BM25, BLAIR-large, and Qwen3-Embedding-8B retrieval instances, while Linq-Embed-Mistral is held out for retriever-generalization evaluation.Per category, data are split 75:15:10 into train, dev, and test; eligible query-retriever pairs require the positive item in the top 100.Source: §3.3 · p. 4
• Each eligible pair yields sampled five-way instances with negatives from three rank buckets, exposing training to different portions of the top-100 candidate distribution.Two negatives are sampled from ranks 1–20, one from 21–50, and one from 51–100; eight instances are used for training and five for development.Source: §3.3 · p. 4
• GRPO optimizes extractor memories using downstream positive-selection accuracy on five-candidate sets, supplemented by deterministic checks for concise, well-formed, supported memory.The checks encourage non-empty within-category memory and penalize malformed tags, repetition, overlong output, placeholders, and unsupported evidence-like content; they regularize rather than replace reranking utility.Source: §3.3 · p. 4

4.1 Baselines

• Baselines include no memory, raw purchase history, GPT-5.5-extracted memory, and MEMRERANK without RL; MEMRERANK uses an RL-trained extractor for query-independent preferences.No-memory reranking receives only the query and candidate products, while the raw-history baseline appends purchase-history products directly to the prompt.Source: §4.1 · pp. 4–5
• MR.Rec and Mem0 are run with Qwen2.5-7B-Instruct and the same metadata-only history window, with their outputs rendered into the reranker’s memory fields.MR.Rec summarizes category-level preferences into a profile, whereas Mem0 extracts atomic preference facts.Source: §4.1 · p. 5

4.2 Implementation Details

• Qwen2.5-7B-Instruct is the memory-extractor backbone; category-specific memory targets and checkpoints are selected using development-set reranking performance.The main table uses within+cross memory for Electronics and within-category memory for Beauty & Personal Care.Source: §4.2 · p. 5
• Training uses five epochs, rollout size 8, and o4-mini as reward reranker; utility is the fraction of sampled responses selecting the positive product.The reward reranker is sampled five times per training instance.Source: §4.2 · p. 5
• Final top-100 setwise evaluation uses o4-mini and reports MRR@10 as the primary metric, with candidate and history metadata capped at 120 words.All methods use temperature 1.0, top_p 1.0, and reasoning_effort=low; evidence fields are removed before reward scoring and final reranking, and MRR@5 is reported in Appendix A.1.Source: §4.2 · p. 5

5.1 Main Results

• Across categories, MEMRERANK raises macro-average MRR@10 from 7.41 to 8.00, with category averages rising from 10.59 to 11.25 in Electronics and 4.33 to 4.75 in Beauty & Personal Care.The gains are consistent at the category-average level, though not for every retriever row.Source: §5.1 · p. 5
• On held-out Linq-Embed-Mistral, MEMRERANK achieves the best MRR@10 in both categories, improving Electronics from 13.07 to 14.36 and Beauty & Personal Care from 5.30 to 6.77 over the strongest baselines.Linq was not used to construct training negatives or select checkpoints.Source: §5.1 · pp. 5–6
• MEMRERANK achieves higher MRR@10 with compact preference memory than raw history and MR.Rec, despite their longer history or memory text.This supports learning a concise intermediate memory rather than exposing the reranker to full purchase history.Source: §5.1 · p. 6, Figure 2 · p. 6

5.2 Ablation Study

• Memory targets are category-dependent: Electronics scores 11.25 with within+cross-category memory versus 10.71 within-category, while Beauty & Personal Care scores 4.75 within-category versus 3.74 within+cross-category.Cross-category history can transfer preferences for Electronics, while unrelated purchases may add noise for Beauty & Personal Care.Source: §5.2 · p. 6, Table 2 · p. 6
• The deterministic memory-quality reward improves average MRR@10 in all four category-target comparisons, with gains of +0.73 and +0.89 on Electronics and +1.40 and +0.51 on Beauty & Personal Care.The quality term is smaller than reranking utility and regularizes memory structure, conciseness, and support without replacing task-level feedback.Source: §5.2 · p. 6, Table 2 · p. 6

6 Conclusion

• MEMRERANK trains a Qwen2.5 memory extractor with downstream setwise reranking feedback and lightweight memory-quality regularization instead of feeding long purchase histories directly to an LLM reranker.The framework produces compact, query-independent preference memory for personalized product reranking.Source: §6 · p. 6
• MEMRERANK achieves the best macro-average MRR@10 across BM25 and dense retrievers on top-100 reranking over Electronics and Beauty & Personal Care, including a held-out dense retriever.The held-out dense retriever was not used for training-negative construction.Source: §6 · p. 6

Limitations

• The evaluation is limited to two Amazon-C4 categories and textual product signals, leaving generality across more categories, multilingual queries, and multimodal signals such as images untested.The passage identifies these broader evaluations as future work.Source: Limitations · p. 7
• Both training reward and final evaluation use the proprietary o4-mini API, which is more expensive at scale than self-hosted open-weight reranking.Future work could explore lower-cost distilled rerankers or hybrid evaluation protocols while retaining the same preference-memory interface.Source: Limitations · p. 7

A.1 MRR@5

• Table 3 reports MRR@5 results for the same test queries as Table 1.Source: Table 3 · p. 10

A.2 Prompts

• The prompt figures cover Mem0 and MR.Rec prompts for extracting preference patterns and user profiles.Source: Figure 3 · p. 10, Figure 4 · p. 11, Figure 5 · p. 11
• Additional figures show prompts for query rewriting and reranking with memory.Source: Figure 6 · p. 11, Figure 7 · p. 12
• A separate prompt figure covers within-category shopping preference memory extraction (M_W).Source: Figure 8 · p. 13

A.3 Dataset

• Cross-category purchase histories are generally longer than within-category histories, motivating separate modeling of within-category and cross-category preference memory.History lengths are compared after excluding the target positive product.Source: Figure 9 · p. 14
• Figure 9’s top plot shows benchmark query counts for each target category, while its bottom plot compares average history lengths within and across categories.Source: Figure 9 · p. 14
• Categories are sorted by benchmark query count, and those with fewer than 100 queries are omitted for readability.Source: Figure 9 · p. 14