acceptodds
Under review as a conference paper at ICLR 2027

MemVista: Domain-Adaptive User Memory Evolution for LLM Recommendation Agents

Abstract

When a compact user memory discards useful history, improving the downstream recommendation policy cannot directly recover that evidence. We study whether memory facets, instructions about which historical distinctions to retain, are a useful target of recommendation learning. ***Which historical distinctions are worth preserving depends on the domain***, even after matching profile token budget and the total size of banks of domain-specific facets. We introduce MemVista to ***learn memory facets automatically from recommendation errors***. Facets configure evidence extraction and profile synthesis, while the recommendation prompt remains fixed and never receives them. Evaluation covers twelve domains and six backbones. With GLM, MemVista outperforms Facets from pooled errors by 4.4 percentage points and Shuffled facets with matched total storage by 5.8 points. Target-aware diagnosis improves over Adaptive facets w/o correctness by 10.7 percentage points and Adaptive facets w/o target by 9.0 points. Frozen facets also improve direct summarization and an existing hierarchical memory manager without further search. A blinded, length-controlled exchange with strong reference memories directly tests the value of the additional retained evidence. Together, these results establish the empirical value of supervised content selection beyond domain priors, memory architecture, and ranking-policy adaptation.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Recommender Systems, Large Language Models, User Modeling, Memory, Personalized Recommendation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_f69wOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MemVista, a framework for domain-adaptive user memory evolution in LLM recommendation agents. The method aims to address the problem that compact user memories may discard useful historical information, which cannot be recovered by downstream recommendation policies. MemVista learns domain-shared memory facets from recommendation errors to guide evidence extraction and profile synthesis, while keeping the recommendation prompt fixed. Experiments on twelve domains and six LLM backbones show that MemVista improves recommendation performance over existing methods, and the learned facets can be transferred to other memory constructors.

Critical strengths
  1. The paper proposes learning domain-shared memory facets from recommendation errors to guide which historical information should be preserved, rather than directly updating user memories.
  2. The paper evaluates MemVista on twelve domains and six LLM backbones, with controlled experiments on facet learning, domain dependence, and transfer to other memory constructors.
Critical weaknesses
  1. Questionable baseline performance and limited evaluation settings. The performance of traditional recommendation baselines is concerning. In Table 1, SASRec achieves an average accuracy of only 0.136, which is much lower than BPR (0.191) and LightGCN (0.272). More importantly, SASRec performs worse than random selection in five domains, including Automotive (0.040) and Musical Instruments (0.067), where random selection among ten candidates has an expected accuracy of 0.100. This is difficult to explain, since SASRec is trained on the large background interaction data rather than only the 300 selected training users. These results raise concerns about the implementation and hyperparameter tuning of traditional baselines. Furthermore, the paper does not provide a public code repository, making it difficult to reproduce and verify these unusual results. Moreover, the main experiments evaluate recommendation by selecting one positive item from ten candidates, with nine negative items uniformly sampled from the same domain. This setting may be too simple to reflect practical recommendation tasks. Although the paper includes a harder 50-candidate evaluation, it does not report the traditional recommendation baselines under this setting. Therefore, the experimental results do not provide sufficient evidence for the claimed advantages over existing recommendation methods.
  2. Insufficient validation of the learned memory facets. The main idea of learning memory facets from recommendation errors is interesting, but the learning process relies heavily on LLM-generated diagnoses and prompt-based revisions. The method assumes that recommendation errors can reveal missing historical information in user memories. However, an incorrect recommendation does not necessarily indicate missing information. It may also result from limitations of the recommendation model or uncertainty in user behavior. The paper evaluates proposed facet sets using validation accuracy, but this does not directly verify whether the generated facets capture the missing information responsible for the recommendation errors.
Questions for authors
  1. Why does SASRec perform worse than random selection in several domains, and can the authors provide training details and additional results to verify its implementation and hyperparameter tuning?
  2. How does MemVista perform under more realistic candidate selection settings, such as popularity-based negative sampling or candidates retrieved by traditional recommendation models?
  3. How can the authors verify that the generated facets capture meaningful historical information missing from recommendation errors, rather than only improving validation accuracy?
10 comments
AuthorsOpenReview ↗

Response to Reviewer f69w (1/3)

Dear Reviewer f69w,

Thank you very much for your interest in our work exploring LLM memory in agentic recommendation, and for recognizing the novelty of learning memory facets from recommendation errors.

"main idea of learning memory facets from recommendation errors is interesting"

We also appreciate your detailed and constructive questions regarding baseline reliability and the validation of learned facets. We address these concerns below with additional experiments and clarifications.

1. SASRec performance and reproducibility

"In Table 1, SASRec achieves an average accuracy of only 0.136, which is much lower than BPR (0.191) and LightGCN (0.272)...These results raise concerns about the implementation and hyperparameter tuning of traditional baselines."[W1]; "Why does SASRec perform worse than random selection in several domains, and can the authors provide training details and additional results to verify its implementation and hyperparameter tuning?"[Q1]

Context from prior benchmarks and task difficulty: Low scores for ID-based/traditional baselines on sparse LLM-based reranking benchmarks are also consistent with recent literature. In MARS (arXiv, 2026), SASRec also trails both BPR and LightGCN in three of four domains under 10-candidate evaluation (Table 2 in MARS). Crucially, MARS evaluates on the instruction-conditioned InstructRec benchmark, where each query includes an explicit user-intent request derived from the target item's review. In contrast, our benchmark removes all target-derived intent instructions and evaluates purely history-only next-interaction prediction, making the task substantially harder even on overlapping domains.


Here, we investigated SASRec's performance by separating evaluation cases according to whether the correct item appeared in the training interactions. An item is considered cold-start if it has no interaction records in the training data.

Item cold start is particularly challenging for SASRec, which primarily relies on item IDs and interaction histories. The model needs to rely on content signals such descriptions, categories, brands, and attributes to make judgments. This situation also arises in practice when newly listed items have no training interactions and item metadata can help rank them.

The table below reports the proportion of evaluation cases with a cold-start target item, overall SASRec accuracy, and accuracy on the cold-start and non-cold-start subsets. Domains are ordered by their cold-start proportions.

DomainCold-start target proportionOverall SASRec accuracyCold-start accuracyNon-cold-start accuracy
Automotive87.0%4.0%0.6%26.5%
Movies & TV82.0%11.3%2.1%53.0%
Games70.3%9.0%0.0%30.4%
Grocery69.0%8.3%1.7%22.9%
Kindle66.3%11.0%1.3%30.2%
Health65.0%10.3%0.0%29.5%
Music58.3%6.7%1.5%14.0%
Yelp58.3%12.7%1.3%28.6%
Baby57.7%13.3%2.3%28.3%
Sports56.0%8.3%2.8%15.3%
Books41.0%19.0%1.3%31.3%
Goodreads25.7%49.0%0.0%66.0%

SASRec exceeds the 10% random-choice accuracy on non-cold-start cases in all twelve domains. In Automotive, 87.0% of evaluation cases have a cold-start target. SASRec achieves 0.6% accuracy on these cases and 26.5% when the target appears in training, yielding 4.0% overall accuracy. In Music, 58.3% of cases have a cold-start target, and accuracy is 1.5% on those cases versus 14.0% when the target appears in training, yielding 6.7% overall accuracy. These results indicate that item cold start contributes substantially to the low overall scores.

"Furthermore, the paper does not provide a public code repository, making it difficult to reproduce and verify these unusual results."[W1]

In fact, before submission, we had also noticed these low scores and reran SASRec separately to check the evaluation pipeline and verify that the results were reproducible. While we initially withheld the codebase due to standard confidentiality considerations during the review stage, our constructive exchange with the reviewer has encouraged us to share it early. We provide the SASRec reproduction code in the supplementary materials, and will release the full training and facet evolution code upon acceptance. We appreciate your request for more implementation detail, which prompted us to report this breakdown.

AuthorsOpenReview ↗

Response to Reviewer f69w (2/3)

2. Candidate selection and popularity-weighted negative sampling

"Moreover, the main experiments evaluate recommendation by selecting one positive item from ten candidates, with nine negative items uniformly sampled from the same domain. This setting may be too simple to reflect practical recommendation tasks."[W1]

As discussed in our response to Reviewer eRqr, “Understanding: shared facets, individualized memories, and research scope (2/2),” the ten-candidate evaluation format is also used in the closely related studies listed in Table R1. MemVista follows this format to facilitate comparison and extends the evaluation from 2–4 domains to 12 domains. To further examine sensitivity to candidate selection, we additionally evaluate popularity-weighted negative sampling below.

Table R1. Domain coverage, data sources, and main candidate setting.

StudyNumber of domainsDomains and data sourcesMain candidate setting
AgentCF (WWW 2024)2Amazon: CDs and Vinyl; Office Products10-candidate evaluation
iAgent / i²Agent (Findings of ACL 2025)4Amazon: Books; Movies and TV. Other sources: Goodreads; Yelp10-candidate evaluation
MemRec (ACL 2026)4Amazon: Books; Movies and TV. Other sources: Goodreads; Yelp10-candidate evaluation
MARS (arXiv, 2026)4Amazon: Books; Movies and TV. Other sources: Goodreads; Yelp10-candidate evaluation
MemVista (Ours)1210 Amazon Reviews 2023 categories, plus Goodreads and Yelp10-candidate evaluation

"How does MemVista perform under more realistic candidate selection settings, such as popularity-based negative sampling or candidates retrieved by traditional recommendation models?"[Q2]

To examine sensitivity to negative sampling, we replace uniform sampling with popularity-weighted sampling while retaining the true next item and nine negatives (ten candidates in total). Negatives are sampled without replacement within each domain, with weights proportional to (c_i + 1)^0.25, where c_i is the item's interaction count in the original training data.

This weighting increases the probability of sampling popular items. The results are shown below.

DomainUniform samplingPopularity-weighted sampling
Goodreads0.670.69
Yelp0.530.57
Movies & TV0.610.59
Music0.640.69
Kindle0.650.72
Baby0.310.34
Automotive0.570.59
Books0.590.57
Games0.490.49
Grocery0.400.41
Health0.280.32
Sports0.380.45
Macro-average0.510.54

MemVista's macro-average accuracy increases from 0.51 to 0.54 under popularity-weighted sampling. Accuracy improves in nine domains, remains unchanged in one, and decreases in two. These results support its robustness to this change in the negative-sampling distribution.

3. Dependence on LLM-generated diagnoses and prompt-based revisions

“relies heavily on LLM-generated diagnoses”[W2]

Thank you for raising this point. While LLM-based recommendation inherently relies on the capabilities of the underlying backbone, our framework is designed to be model-agnostic and does not depend on a specific proprietary model. To verify this robustness, we evaluated our method across six different open- and closed-source models in Table 1, demonstrating consistent performance gains across varying model capacities.

“and prompt-based revisions”[W2]

We agree with this observation. Our key extraction and synthesis operations are indeed guided by natural-language prompts, and the evolved domain-shared facets (Pd\mathcal{P}_d) are stored as textual content.

Rather than relying on brittle prompt engineering, however, this design directly drives our generalization. By keeping all model weights, the memory architecture, and the ranking prompt completely fixed (the ranker never sees any facet text), we optimize what historical distinctions to retain via weight-frozen, error-driven evolution in textual space. Three results confirm that the evolved facets capture generalizable domain knowledge rather than prompt-specific artifacts:

  1. Outperforming static priors (Table 2): +10.3 pp over hand-crafted Fixed domain-prior facets (0.510 vs. 0.407).
  2. Cross-constructor transferability (Table 3): Once frozen, the learned facets transfer directly to completely different memory architectures—such as single-call Direct history summary (+12.3 pp) and hierarchical MARS (+11.8 pp)—without any re-tuning or prompt adaptation.
  3. Cross-backbone consistency (Table 1): The exact same evolution procedure consistently yields the best macro accuracy across all six open- and closed-source backbones.
AuthorsOpenReview ↗

Response to Reviewer f69w (3/3)

4. Whether learned facets capture meaningful missing information

“However, an incorrect recommendation does not necessarily indicate missing information. It may also result from limitations of the recommendation model or uncertainty in user behavior...The paper evaluates proposed facet sets using validation accuracy, but this does not directly verify whether the generated facets capture the missing information responsible for the recommendation errors.” [W2]

We agree that errors have multiple causes and that validation accuracy alone cannot confirm that an error stems from missing information. MemVista uses errors as feedback for proposing and revising facets, with validation performance determining whether the proposed revisions are retained, rather than assuming that every error reflects missing information.

While inherent LLM limitations and behavioral uncertainties cannot be eliminated, MemVista explicitly accounts for them by requiring extracted evidence to preserve uncertainty (Section 3.2, Appendix C) and evaluating the effectiveness / stability of facet revisions across three random seeds and six backbones (Tables 1–2, Figure 3).

Importantly, our evidence goes beyond validation accuracy: we examine what information the facets preserve and directly test whether the additional facts improve recommendation.


What information is preserved? As discussed in our response to Reviewer eRqr, "Understanding about domain facets," Appendix G analyzes facet composition, evolution, and associated held-out gains across all twelve domains (Figures 5–8). Figure 9 provides conceptual examples illustrating the distinctions these facets help preserve, including routine context, changing needs, and series progress. Appendix I provides concrete facets that retain relationships beyond broad preferences—for example, component roles and purchase order within a vehicle repair project, or dependencies between gaming hardware and accessories.

We apologize for not making these analyses more prominent given the limited space. We placed the detailed facet analyses in the appendix and prioritized the main text for comparisons with existing methods, which we consider the clearest evidence of our approach’s effectiveness.

Does the retained information help? Section 4.6 and Appendix E directly test whether historical facts preserved by MemVista—but omitted from reference profiles—improve recommendation. On 600 test users, 50 per domain, we replace at most two facts in each eligible reference profile with history-supported facts retained only by MemVista. We then repeat recommendation with the same ranker and candidates.

To isolate the value of the information, we standardize the profiles’ wording and match fact count and token length. The exchanged facts come from older history, while recent context and decoding remain fixed. Users are selected independently of prediction outcomes, and annotators verify facts without seeing candidates, targets, predictions, or method identities.

Replacing reference facts with MemVista-retained facts improves all three reference methods (Table 4 in MemVista):

Reference memoryBefore exchangeAfter exchangeGain (pp)Paired 95% interval (pp)
Adaptive facets w/o correctness0.4070.425+1.8[0.7, 3.2]
Adaptive facets w/o target0.4170.433+1.7[0.5, 2.8]
Independent facet proposals0.4930.508+1.5[0.3, 2.7]

Both columns use the standardized, length-matched profiles. All 600 users remain in the evaluation, including those without an eligible exchange. All three gain intervals are above zero. These results provide direct evidence that learned facets preserve useful historical information: transferring that information improves recommendation without increasing profile length or changing the ranking model.

The reverse exchange supports the same conclusion (§4.6, p. 9, lines 434–437; Appendix E). Replacing these additional historical facts in MemVista profiles reduces recommendation accuracy more than replacing other facts of similar age and length. This further supports the usefulness of the information preserved by the learned facets.

Together, these results support learning useful facets from imperfect error feedback. Validation and independent testing assess whether the proposed revisions remain useful beyond the diagnosed cases. The content analyses and controlled exchanges further show that the resulting profiles retain additional, history-supported facts that contribute to recommendation performance.


We hope these explanations resolve your concern about learning meaningful facets from imperfect error feedback and strengthen your confidence in the effectiveness of our approach for LLM-based recommendation.

Reviewer_Y7sROpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes MemVista, a domain-adaptive memory framework for LLM recommendation agents. Instead of relying on predefined memory structures, MemVista learns domain-specific memory facets from recommendation errors to guide evidence extraction and user profile construction. Experiments across twelve recommendation domains and six LLM backbones demonstrate improvements over existing recommendation and memory-based approaches.

Critical strengths
  • The paper presents an interesting approach to learning what information should be preserved in user memory based on downstream recommendation feedback.
  • The proposed domain-adaptive memory evolution mechanism is well motivated and demonstrates meaningful improvements over fixed memory configurations.
  • The experiments cover diverse domains and LLM backbones, with comprehensive ablations and controlled comparisons supporting the effectiveness of the learned memory facets.
Critical weaknesses
  • The proposed method primarily relies on iterative prompt optimization and error-driven refinement, making its methodological novelty relative to existing reflective optimization approaches somewhat limited.
  • The evaluation primarily uses ten-candidate recommendation tasks with randomly sampled negatives, which may not adequately reflect realistic recommendation challenges.
  • The iterative facet optimization process requires substantial offline computation, but its computational cost and scalability are not sufficiently evaluated.
  • The framework assumes fixed domain-specific facets and does not investigate their effectiveness under evolving user preferences or dynamically changing domains.
Questions for authors
  • How does MemVista compare with general-purpose prompt optimization methods such as GEPA under strictly matched computational budgets and repeated runs?
  • How well does MemVista perform in more realistic recommendation settings involving larger candidate pools and harder negative samples?
  • What is the computational cost of facet evolution, and how does it scale with the number of domains and users? Can the learned memory facets adapt efficiently to changing user preferences or transfer to previously unseen recommendation domains?
5 comments
AuthorsOpenReview ↗

Appreciating the opportunity to discuss our work with Reviewer Y7sR

Dear Reviewer Y7sR,

Thank you for recognizing the motivation of learning what information to preserve in user memory, the improvements over fixed memory configurations, and the breadth of our controlled experiments! We address your questions about evaluation settings, computational cost, transfer, and the GEPA comparison below.

AuthorsOpenReview ↗

Response to Reviewer Y7sR (1/3)

1. Candidate selection and popularity-weighted negative sampling

"The evaluation primarily uses ten-candidate recommendation tasks with randomly sampled negatives, which may not adequately reflect realistic recommendation challenges."[W2]

The ten-candidate evaluation format is established in closely related agentic recommendation studies. As Table R1 shows, these studies evaluate on 2–4 domains drawn from Amazon or a combination of Amazon, Goodreads, and Yelp. MemVista follows this format and extends the evaluation to 12 domains. We complement this common protocol with popularity-weighted negatives and the larger Hard-50 candidate setting below.

Table R1. Domain coverage, data sources, and main candidate setting.

StudyNumber of domainsDomains and data sourcesMain candidate setting
AgentCF (WWW 2024)2Amazon: CDs and Vinyl; Office Products10-candidate evaluation
iAgent / i²Agent (Findings of ACL 2025)4Amazon: Books; Movies and TV. Other sources: Goodreads; Yelp10-candidate evaluation
MemRec (ACL 2026)4Amazon: Books; Movies and TV. Other sources: Goodreads; Yelp10-candidate evaluation
MARS (arXiv, 2026)4Amazon: Books; Movies and TV. Other sources: Goodreads; Yelp10-candidate evaluation
MemVista (Ours)1210 Amazon Reviews 2023 categories, plus Goodreads and Yelp10-candidate evaluation

"How well does MemVista perform in more realistic recommendation settings involving larger candidate pools and harder negative samples?"[Q2]

Popularity-weighted negatives.We replace uniform sampling with popularity-weighted sampling while retaining the true next item and nine negatives (ten candidates in total). Negatives are sampled without replacement within each domain, with weights proportional to (c_i + 1)^0.25, where c_i is the item's interaction count in the original training data. This increases the probability of sampling popular items.

Table R2. GLM accuracy under different negative-sampling distributions.

DomainUniform samplingPopularity-weighted sampling
Goodreads0.670.69
Yelp0.530.57
Movies & TV0.610.59
Music0.640.69
Kindle0.650.72
Baby0.310.34
Automotive0.570.59
Books0.590.57
Games0.490.49
Grocery0.400.41
Health0.280.32
Sports0.380.45
Macro-average0.510.54

MemVista's macro-average accuracy increases from 0.51 to 0.54 under popularity-weighted sampling. Accuracy improves in nine domains, remains unchanged in one, and decreases in two. These results support robustness to this change in the negative-sampling distribution.

Larger candidate pools with semantically similar negatives. The submitted paper also includes Hard-50 (Appendix E, Table 9): one target and 49 negatives sampled from its 200 nearest eligible items in a frozen text-embedding space. All methods retain their original selected configurations, without retuning on Hard-50. This is a target-neighbor stress test.

Table R3. Hard-50 results reported in the submitted paper.

MethodHard-50 accuracy
No facets0.235
MemRec0.270
Facets from pooled errors0.298
Independent facet proposals0.310
MemVista0.329

MemVista retains the highest accuracy among these methods when the candidate pool expands to 50 items with similar negatives, exceeding MemRec by 5.9 percentage points.

AuthorsOpenReview ↗

Response to Reviewer Y7sR (2/3)

2. Offline computation and scalability

"The iterative facet optimization process requires substantial offline computation, but its computational cost and scalability are not sufficiently evaluated."[W3]

"What is the computational cost of facet evolution, and how does it scale with the number of domains and users?"[Q3]

Facet evolution is performed offline at the domain level, and the resulting facets are reused across users. We report GLM training-token costs below; k denotes one thousand tokens. The 300-user, one-domain configuration uses 1,930k input tokens and 876k output tokens, totaling 2,806k.

Table R4. GLM offline facet-evolution token cost.

StageInput tokens (k)Output tokens (k)Total tokens (k)
Offline facet evolution: 300 training and 300 validation users, one domain1,9308762,806

Table R5 shows that training-token consumption grows approximately linearly with the number of domains and increases more gradually with the number of users per split.

Table R5. Total GLM training-token consumption as user and domain counts increase (k).

Users per training/validation split per domain1 domain3 domains5 domains
3002,8068,12113,050
4003,4699,23816,957
5003,82310,74319,671

Each row uses the indicated number of training users and, separately, validation users in every domain. Each cell sums input and output tokens across all domains in that configuration.

Once learned, each domain’s facets are shared across users, so adding users requires only constructing their individual memories.

Reviewer_eRqrOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

MemVista tackles next-item recommendation with a fixed LLM ranker: the ranker sees a short profile (at most 256 tokens) plus the last 5 interactions and must pick the user's actual next item from 10 candidates. The paper's point is that profile content, not ranker strength, is the bottleneck: the distinctions worth preserving are domain-level and shared across users in a domain. It therefore learns, per domain, a small set of "facets" — natural-language instructions telling a constructor what distinctions to extract and keep, like items on a diagnostic questionnaire. The set is learned offline from training errors (diagnose 4 errors, rewrite the set, keep only strict validation improvements, up to 12 rounds) and frozen before testing. Headline result: a 5.2–10.3 point macro gain over the strongest baseline across 6 backbones and 12 domains. The question is whether this amounts to a sufficient contribution, given how narrow the demonstrated setting is.

Critical strengths
  • The domain-shared facet set is a clean idea: one diagnostic questionnaire per domain, instead of per-user online memory or a generic "summarize this user" prompt.
  • The facets are discovered, not hand-written: the loop finds the diagnostic questions from its own errors, with validation as the gate, and the gains hold across backbones within the tested setting.
Critical weaknesses
  • The contribution does not go beyond a good idea. What is learned is the content-selection rule, but the loop itself is a standard diagnose–propose–validate cycle, and the idea is not pushed deeper: there is no per-user refinement on top of the shared list, no cross-domain user modeling, and no analysis of what the learned facets actually capture beyond aggregate accuracy. The paper demonstrates the idea in one narrow setup and stops there; as a contribution, that is too thin.
  • The demonstrated setting is far from deployment. Test users are disjoint per domain and facets are learned independently per domain, so the architecture has no cross-domain user identity: a real user who reads books and buys games has one taste spanning domains, but here the two profiles are unrelated, and the cross-domain study transfers facet sets, not users. The task is also 10 uniform negatives rather than retrieval, Hard-50 negatives are target-neighbor constructions rather than deployment candidates, the cold-item rate is zero by construction, and facets are frozen with no drift or long-term analysis.
  • The experiments leave the central claims untested. Learned from majority errors and capped at 12 items, with an aggregate validation gate, a facet helping a minority a lot while slightly hurting the majority can be rejected on total score — yet only domain averages are reported, with no per-user gain distribution and no subgroup analysis. The obvious hybrid (domain list plus per-user refinement) is never ablated, and there is no per-facet attribution: only whole sets are selected. Each Table 1 cell rests on 300 users per domain, the three runs reuse the same users, and the 68/72 headline is reported as point estimates with no intervals; the same 300 validation users are reused for 12 rounds of selection, yet no validation accuracy, no validation-to-test gap, no second held-out set, and no acceptance counts are reported. Two knobs are never varied: the 4-errors-per-round count and the 256-token budget (the sweep reuses facets learned at 256, with no accuracy-cost-latency curve).
Questions for authors
  1. Can you report the per-user gain distribution and subgroup breakdowns (user type, history length, taste rarity), especially which users get worse and how minority types fare?
  2. What happens with a hybrid — domain list plus per-user refinement? Does personalization add anything once the shared list is in place?
  3. How would cross-domain user modeling work: where is a single user's cross-domain taste stored and transferred?
  4. Can you add intervals to Table 1, report validation accuracy and the validation-to-test gap, and ideally a second held-out validation set?
  5. What is the sensitivity to the diagnosis count (4 vs 8 or 16 errors per round) and to the budget when facets are re-learned at each budget, with an accuracy-cost-latency comparison?
8 comments
AuthorsOpenReview ↗

Appreciating the opportunity to discuss our work with Reviewer eRqr

Hi, eRqr. Thank you for the care you put into reviewing our work and for your interest in our paper, which explores the use of LLM memory in recommendation. We also appreciate your thoughtful insights.

Our response has two parts. The first revisits our approach and clarifies several points that may have been overlooked, with the hope of reaching a shared understanding. The second focuses on experiments that address your questions about the paper and, in particular, your interest in how our method works for different user groups.

AuthorsOpenReview ↗

Understanding: shared facets, individualized memories, and research scope (1/2)

Understanding about domain facets:

"no per-user refinement on top of the shared list"[W1]; "Learned from majority errors and capped at 12 items"; "a facet helping a minority a lot while slightly hurting the majority can be rejected on total score...The obvious hybrid (domain list plus per-user refinement) is never ablated.[W3]"

Our work builds on the premise that users within a domain can share decision-relevant patterns even when their specific preferences differ. We learn these patterns from feedback across users and distill them into compact, reusable memory facets that guide individualized memory construction.

The key distinction we hope to clarify is between individual preferences and shared patterns that help interpret them. What is shared is what to look for in a user’s history; what is personalized is the evidence retained from that history. For example, a facet about series progression guides the system to retain which series a reader follows and how far they have progressed. These details differ across readers: the question is shared; the answer is user-specific. Note that: Shared facets therefore support personalized profiles by capturing domain-level regularities while preserving each user’s distinct history and preferences. Appendix F gives the illustration.

Your question about per-user refinement raises a further issue: how can shared domain knowledge and individual reflection complement each other? To examine this concretely, we provide a case from the prediction logs below. User reflection captures the user’s individual preferences, while the domain facet adds a consideration that reflection alone overlooked: purchasing another edition of a familiar title may represent a collection upgrade.

ItemContent
1. User reflection onlyThe user enjoys crime dramas featuring moral conflicts and complex characters, as well as historical war dramas based on true events. They value production quality and acting.
2. Incorrect choice without facets and its rationaleThe Sopranos: Season 6, Part 2 (ID 7234; incorrect). The original output judged it consistent with the user’s preferences for crime dramas, moral complexity, and strong performances. It explicitly rejected Band of Brothers, describing it as merely a repackaged edition of something the user had “already watched.”
3. Domain facetRepurchases and edition upgrades in film and television collections (H8v3): Collectors may purchase different editions of familiar titles. Substantial overlap between a candidate and previously acquired titles can signal collection completion or an edition upgrade. Title overlap should therefore be considered a potentially positive signal rather than automatically dismissed as redundant.
4. Correct choice with the facet and its rationaleBand of Brothers: Blu-ray Reissue (ID 311; correct). Historical record R2 shows a previous purchase of Band of Brothers on DVD, while the candidate is a Blu-ray edition of the same title. With the facet, the model interpreted this relationship as a collection upgrade and selected the actual next purchase. The title also matched the user’s preference for historical war dramas.

Here, the domain facet complements the user’s preference for historical war dramas by helping interpret the specific DVD purchase in their history.


"and no analysis of what the learned facets actually capture beyond aggregate accuracy"[W1];"and there is no per-facet attribution: only whole sets are selected."[W3]

We apologize for not making these analyses more prominent. Appendix G (Evolution and Composition Analyses) examines what the learned facets capture across all twelve domains beyond aggregate accuracy: Figure 5 reports their semantic composition and associated held-out gains; Figure 6 tracks the composition of remaining errors during evolution; and Figures 7–8 show facet trajectories and category-level activations and net gains.

Figure 9 illustrates how facets help preserve previously overlooked historical distinctions, including routine context, changing needs, and series progress. Appendix I provides concrete facet examples specifying which relationships in user histories to retain.

Given the limited space, we placed the detailed facet analyses in the appendix and prioritized the main text for comparisons with existing methods, which we consider the clearest evidence of our approach’s effectiveness.

AuthorsOpenReview ↗

Understanding: shared facets, individualized memories, and research scope (2/2)

Understanding about user modeling:

"no cross-domain user modeling”[W1]---- the cross-domain study transfers facet sets, not users"[W2]; "How would cross-domain user modeling work: where is a single user's cross-domain taste stored and transferred?"[Q3]

Scope of user modeling and domain-wise evaluation. To answer W2 directly: the evaluated system does not maintain a joint cross-domain memory for the same user. The cross-domain experiment transfers the rules used to build memories (the facet sets), not an identified user’s preferences. MemVista does construct individualized user memories, but it does so within each domain. This is by design (W1).

For Q3: we do not store or transfer a single user’s taste across domains. In this setting it also reduces to self-reflection, rather than testing domain-wise memory construction, A joint cross-domain profile is outside the problem MemVista studies.

Domain-wise evaluation is an established setting in agentic recommendation. As Table R1 shows, closely related studies evaluate on 2–4 domains drawn from Amazon or a combination of Amazon, Goodreads, and Yelp. MemVista follows the same ten-candidate format and extends the evaluation to 12 domains: ten Amazon Reviews 2023 categories, Goodreads, and Yelp.

Table R1. Domain coverage, data sources, and main candidate setting.

StudyNumber of domainsDomains and data sourcesMain candidate setting
AgentCF (WWW 2024)2Amazon: CDs and Vinyl; Office Products10-candidate evaluation
iAgent / i²Agent (Findings of ACL 2025)4Amazon: Books; Movies and TV. Other sources: Goodreads; Yelp10-candidate evaluation
MemRec (ACL 2026)4Amazon: Books; Movies and TV. Other sources: Goodreads; Yelp10-candidate evaluation
MARS (arXiv, 2026)4Amazon: Books; Movies and TV. Other sources: Goodreads; Yelp10-candidate evaluation
MemVista (Ours)1210 Amazon Reviews 2023 categories, plus Goodreads and Yelp10-candidate evaluation

Advances in recommendation memory can be established within domain-wise evaluation. MemRec's collaborative memory and MARS's hierarchical memory exemplify this. MemVista investigates learning what evidence to retain from recommendation feedback: with the constructor, ranker, and memory budgets fixed, learned facets improve twelve-domain mean accuracy by 10.3 percentage points over fixed domain-prior facets (GLM-5.2, Table 2 in the paper). This directly tests memory-content learning; modeling the same user across domains additionally requires cross-domain preference fusion.

We agree that jointly modeling the same user across domains is valuable, but it addresses a different problem from ours. It may require integrating the user’s histories across domains and determining how much weight to give each domain and how their preferences interact. MemVista focuses specifically on learning which distinctions within a domain deserve attention when constructing individual user memories. We will explicitly clarify this scope and distinguish it from joint cross-domain user modeling in the revised paper.

"10 uniform negatives rather than retrieval"..."the cold-item rate is zero by construction, and facets are frozen with no drift or long-term analysis."[W2]

Our evaluation follows the domain-wise, 10-candidate next-item protocol used in related work (Table R).

MemVista’s facets guide which distinctions to preserve from a user’s history within a domain; they do not encode item identities or item-specific features. The method therefore focuses on user-memory construction, while our evaluation does not separately measure performance on cold items (follows a widely used benchmark protocol in Table R). We will clarify this scope.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.