acceptodds
Under review as a conference paper at ICLR 2027

HabitusMem: Bayesian Belief Memory over Evolving User Preferences for LLM Agents

Abstract

Personalized LLM agents must act on a user's current preferences, which evolve across interactions and are rarely stated in full. Existing memory systems store and retrieve content from the interaction history, leaving the agent to work out from fragments which preferences still hold and where they apply. Drawing on Bourdieu's account of habitus as durable but revisable dispositions, we model the user as a latent preference state that every interaction, including the current request, both reveals and updates. Bayesian inference over this state yields the assistant's belief about the user, and Bayesian decision theory specifies what memory should convey: the part of the belief that bears on the agent's current decision. We implement this view in HabitusMem, a supervised writer that renders the belief as text records for a frozen agent, each stating a claim with its evidence, scope, and status. The writer's training tasks mirror belief inference, belief update, and decision projection, and one writer serves both personalized question answering and personalized agent tasks. Experiments on PersonaMem-v2 and VitaBench 2.0 with two agent families show that HabitusMem outperforms existing memory systems in personalized question answering and yields the most reliable task completion for personalized agents among all non-privileged conditions, surpassing even full-history access.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Personalization, LLM Agent, Memory

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_uD1cOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes HABITUSMEM to address the problem of long-term user memory in personalized agents. The core argument is that, rather than treating memory as a compression of historical conversations, it should be viewed as a belief about the user’s latent state, with the system deciding what information to retain, update, or retract based on the current request. The authors further introduce structured fields such as claim, evidence, scope, and status, and train a memory writer through supervised fine-tuning. The method is evaluated on PersonaMem-v2 and VitaBench 2.0.

Critical strengths
  1. The paper does not simply treat personalized memory as a compression of historical information. Instead, it emphasizes that memory should serve the current user request and attempts to distinguish among claim, evidence, scope, and status. This structured representation is intuitively useful for handling user preferences, changing user states, and potentially conflicting information, and it aligns well with the practical requirements of long-term memory agents.
  2. The paper analyzes the effects of static writer, current request conditioning, and different training tasks, and separately investigates forgetting/retraction scenarios. These experiments provide some discussion of the method’s applicability and failure modes.
Critical weaknesses
  1. There is a clear gap between the Bayesian formulation and the actual algorithm. The paper proposes a Bayesian belief and a risk-based objective, but the actual implementation does not estimate the posterior or explicitly optimize this objective. Instead, it trains a teacher-supervised memory writer. Therefore, it is currently unclear whether the Bayesian formulation itself contributes to the performance gains.
  2. The contribution of teacher-generated supervision is difficult to disentangle. The training targets are primarily generated by a teacher, and their factual correctness is not systematically verified. Therefore, the observed performance gains may primarily come from teacher distillation, request conditioning, or the structured schema, rather than from the proposed Bayesian principle itself.
  3. The forgetting/retraction capability remains relatively weak. On forgetting queries, HABITUSMEM significantly underperforms full context and some baselines, and it performs particularly poorly on active near-duplicate contradictions. This suggests that the proposed status/retraction representation does not yet adequately resolve the problems of stale memories and contradictions.
Questions for authors
  1. If the Bayesian formulation were completely removed while keeping the same backbone, training data, teacher targets, request conditioning, and memory schema, how would the performance of HABITUSMEM change?
  2. How is the quality of the teacher-generated targets verified? The training targets in the paper are primarily produced through teacher generation rather than being subjected to human or automated factual-consistency verification. Has the authors’ evaluation checked for hallucinations, incorrect preferences, incorrect scopes, or incorrect retractions in the teacher-generated memories? If the teacher targets contain systematic errors, would the student model simply inherit those errors?
8 comments
AuthorsOpenReview ↗

Response to Reviewer uD1c (1/2)

Weakness 1 and Question 1: Contribution of the Bayesian formulation

We agree that the implementation does not estimate the posterior or optimize the risk directly. The revision removes "Bayesian" from the title and presents the formulation as the design rationale (General Response, Part 2).

Question 1. The formulation is not a module of the trained system. If it were removed while the backbone, training data, teacher targets, request conditioning, and memory schema all stayed the same, the trained writer and its performance would be unchanged. The formulation contributes by motivating two of the components the question holds fixed, request conditioning and the schema, and Tables G1 and G2 of the General Response ablate each separately. We agree with the reviewer that the gains come from these components rather than from Bayesian inference, and the revision states this.

Weakness 2: Teacher distillation, request conditioning, and schema

The revision separates these three factors (Tables 4 and 5).

Request conditioning and schema. Removing the request from the trained writer at inference lowers accuracy by 13.7 and 14.5 points (Table G1). Removing the record fields and retraining end to end lowers it by 1.3 and 1.9 points (Table G2).

Teacher distillation. The training lineage separates training on teacher-written memories from learning to use the request (PersonaMem-v2, Table 5):

WriterTrained on teacher-written memoriesReceives the requestQwen Acc.DeepSeek Acc.
Base, staticNoNo0.3830.329
Base + requestNoYes0.3930.345
Trained, staticYesNo0.4060.335
Parent writerYesYes, trained to use it0.5050.457
HabitusMemYesYes, trained to use it0.5370.484

Training on teacher-written memories without the request, or adding the request without training, each raises accuracy by at most 2.3 points. The large gain appears only when training teaches the writer to use the request. Distillation from the teacher alone therefore does not explain the result.

AuthorsOpenReview ↗

Response to Reviewer uD1c (2/2)

Weakness 3: Forgetting and retraction

Our claim is narrower than complete forgetting: an explicit status makes a retraction representable, auditable, and repairable, but it is not by itself sufficient for a frozen agent to obey it. The evidence shows that the status representation works and that the remaining errors arise where the agent consumes it.

The writer records the retraction.

CheckResult
PersonaMem-v2 ask_to_forget queries whose memory contains a status: retracted record99.3% of 1,048
Controlled probes whose memory keeps the retraction10/10
Controlled retraction probes answered correctly, Qwen1/10
Controlled retraction probes answered correctly, DeepSeek2/10

The retraction is present, yet the agent often does not act on it. This is a memory–agent interface problem, not evidence that the status is absent or meaningless.

The remaining errors concentrate in two interface patterns:

PatternFrequencyAccuracy with the patternAccuracy without it
The retraction names only an abstract topic, not the revoked item85.7% of retraction records0.4530.740
An active near-duplicate of the revoked preference remains12.4% of memories0.2620.526

The frozen agent is sensitive to the precision of the record: it often fails to apply an abstract retraction, and a stale active duplicate can override the retracted one.

The interface is repairable without retraining the agent. We edited the stored memories and re-answered the same queries with the same agents:

MemoryAll forgetting queries, QwenAll forgetting queries, DeepSeekActive near-duplicate, QwenActive near-duplicate, DeepSeek
Stored memory0.4660.2870.1220.122
+ revoked item named0.7350.4990.5370.195
+ conflicting active record removed0.7410.5200.8780.561

The same agents honor a precise retraction, and the status and scope fields make the repair a local edit to the record.

Question 2: Quality of teacher-generated targets

Three properties of the pipeline limit how much the result depends on teacher errors.

  1. The decisive record comes from annotations. In every direct-writing target, the record for the preference that a request depends on is taken from the benchmark's annotation, including whether it is active, superseded, or retracted. The remaining records of the target come from the cold-start writer, which was trained on teacher memories, and the auxiliary targets come from the parent writer.
  2. The evaluation is end to end on held-out personas. Errors inherited from the teacher would lower accuracy on the 5,000 test queries. Training on teacher-written memories alone (Trained, static) gives 0.406 and 0.335, close to the untrained writer, so the gains do not come from reproducing teacher memories (Weakness 2 above).
  3. The writer's outputs are checked on the fields the reviewer lists. In the controlled probes (App. D.1), manual inspection confirms that all ten gift memories attribute the item to the gift recipient rather than to the user, and the full writer records the conditional scope and the retraction in all ten memories and the supersession in nine or ten.

The probes and the end-to-end results indicate that teacher errors do not dominate the writer's behavior on these benchmarks.

We hope that our responses have addressed your concerns, and we would be grateful if you would consider raising your score.

AuthorsOpenReview ↗

Followup on Rebuttal

Dear Reviewer;

Dear Reviewer; Thank you again for taking the time to review our responses. As the rebuttal period will be ending soon, we would greatly appreciate it if you could kindly revisit our responses and let us know whether any concerns or questions remain unaddressed.

We would be more than happy to provide further clarification or additional explanations if anything is still unclear. If our responses and the additional experiments have adequately addressed your concerns, we would also sincerely appreciate your reconsidering the overall score in light of the new evidence. Thank you again for your time and consideration.

Thanks

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

The paper proposes HabitusMem, a “belief memory” for personalized LLM agents that reframes memory as the assistant’s current belief over an evolving latent user state and specifies what memory should convey via Bayesian decision theory. Experiments on PersonaMem-v2 and VitaBench 2.0 with two agent families show that HABITUSMEM outperforms existing memory systems.

Critical strengths
  1. The paper addresses an important and practically relevant problem. It effectively addresses the common pitfall of traditional memory systems conflating one-off contextual instructions with durable long-term preferences.

  2. The experimental evaluation is comprehensive, and the reported improvements over strong baselines are encouraging.

  3. The paper is well organized and easy to follow.

Critical weaknesses
  1. There is a significant disconnect between the heavy Bayesian theoretical framing and the actual system implementation. No explicit posterior is estimated, and decision risk is not optimized end-to-end.

  2. The history selection policy is heuristic and may exclude decisive late evidence.

  3. Memory is recomputed per request from a selected raw log rather than incrementally maintained as a persistent posterior state, which partially undercuts the “belief tracking” claim.

  4. It is unclear if the baselines received comparable task-specific adaptation, for HabitusMem is SFT-trained whereas some baselines are primarily prompting-based.

  5. The experiments do not include the most relevant or recent state-of-the-art baselines [1,2,3].

  6. The paper lacks a thorough analysis of scalability and deployment costs.

  7. The experiments lack statistical significance testing.

  8. No source code is provided, making the results difficult to reproduce.

[1] Yu H, Chen T, Feng J, et al. Memagent: Reshaping long-context llm with multi-conv rl-based memory agent[C]//International Conference on Learning Representations. 2026, 2026: 39458-39486.

[2] Wang Y, Takanobu R, Liang Z, et al. Mem-{\alpha}: Learning memory construction via reinforcement learning[J]. arXiv preprint arXiv:2509.25911, 2025.

[3] Shen X, Zhou Y, Wu Y, et al. Agentic Recommender System with Hierarchical Belief-State Memory[J]. arXiv preprint arXiv:2605.14401, 2026.

Questions for authors

See Weaknesses.

8 comments
AuthorsOpenReview ↗

Response to Reviewer WKGL (1/2)

Weakness 1: Bayesian framing versus implementation

We agree. No posterior is estimated and the decision risk is not optimized end to end, as Section 3.1 states. The revision removes "Bayesian" from the title and presents Equations 1–2 as the design rationale for what the memory should state. Part 2 of the General Response lists the three design choices that follow from this rationale and the controlled comparisons that test each of them (Tables G1–G3).

Weakness 2: History selection

Thank you for your comment. We retain early history to preserve user background and longstanding preferences. On VitaBench, we additionally retain recent interactions to capture current execution constraints. Whenever the history or current request changes, the writer infers, selects, and regenerates the memory, allowing it to adapt to new evidence and the current question rather than remaining a fixed summary. The history selector determines which evidence is visible. The writer then constructs a request-specific memory, identifying relevant beliefs and marking superseded ones.

We add an experiment on PersonaMem-v2 using DeepSeek, with the same writer and input/output caps for both selectors.

History selectorAccuracy
Ours48.42%
Head-and-tail, with 15% prefix and 85% recent history46.86%

These results show that our selection strategy outperforms the recency-heavy head-and-tail selector.

Weakness 3: Recomputation per request

Thank you for your comment. HabitusMem is trained as a shared belief-state model: direct-writing examples teach belief inference, state-update examples teach how new evidence should revise earlier claims, and projection examples teach which beliefs matter for the current request. At inference, the writer applies this learned mapping to the selected history and request, reconstructing a memory that identifies relevant beliefs, their supporting evidence and applicability, and whether they remain active, have been superseded, or have been retracted. Reconstructing from raw evidence also allows earlier interpretations to be reconsidered without relying solely on a previously compressed memory. Thus, our implementation supports belief reconstruction and revision through a trained writer, while not maintaining a persistent posterior through recursive online updates. We will clarify this meaning of "belief tracking."

Weakness 4: Comparable adaptation of baselines

Thank you for your comment. We described the experimental setup in Section 4.1. We evaluate all methods under one protocol: fixed model parameters, the same downstream agents, test tasks, tool environments, and scoring procedure. VitaBench's shared personalization prompts are part of that protocol, and we do no additional parameter training on the test tasks. We train HabitusMem's SFT on the official PersonaMem-v2 training split and transfer the same trained writer to VitaBench without further training. The main results therefore compare complete memory systems under one protocol, and our trained controls examine the contribution of training separately.

Weakness 5: More baselines

Thank you for the references. We evaluated MemAgent and Mem-α on both benchmarks.

PersonaMem-v2 accuracy:

MethodDeepSeekQwen
Base LLM (no memory)0.2660.446
Rewrite0.4380.484
A-MEM0.4080.489
MemAgent0.3490.390
Mem-α0.3620.399
HabitusMem0.4840.537

VitaBench 2.0, Qwen agent:

MethodAvg@4Pass@4Pass^4
Base LLM (no memory)0.08950.20100.0182
Rewrite0.31390.61090.0700
A-MEM0.29640.54860.0999
MemAgent0.16700.31520.0558
Mem-α0.17640.33720.0687
HabitusMem0.32910.56160.1310

MemAgent reads a long input in segments and overwrites a fixed-length memory to answer a question about that input, with rewards for answer correctness on long-context question answering. Mem-α learns to construct a general-purpose memory store from rewards on downstream questions. Neither method is trained for personalization. They do not infer a user's preferences from interactions and cannot tell whether a preference still holds or in which situations it applies, so they fail to capture the user's preferences well.

Weakness 6: Scalability and cost

The revision adds these measurements in App. C.3.

AuthorsOpenReview ↗

Response to Reviewer WKGL (2/2)

Weakness 7: Statistical significance

We add statistical significance tests. Each cell gives the gain over the base LLM with its 95% confidence interval; *, **, and *** mark p < 0.05, 0.01, and 0.001, and ns marks a non-significant difference.

PersonaMem-v2 accuracy:

MethodQwenDeepSeek
Full context (reference)+0.123 [+0.105, +0.140] ***+0.261 [+0.244, +0.277] ***
RAG+0.014 [−0.005, +0.033] ns+0.126 [+0.110, +0.142] ***
Rewrite+0.038 [+0.020, +0.055] ***+0.172 [+0.156, +0.188] ***
Mem0+0.024 [+0.008, +0.040] *+0.149 [+0.133, +0.165] ***
A-MEM+0.043 [+0.025, +0.062] ***+0.143 [+0.127, +0.159] ***
MemListener−0.028 [−0.045, −0.012] **+0.110 [+0.095, +0.126] ***
HabitusMem+0.091 [+0.071, +0.111] ***+0.218 [+0.200, +0.236] ***

VitaBench 2.0 average reward (Avg@4):

MethodQwenDeepSeek
Ground-truth profile (privileged)+0.244 [+0.216, +0.272] ***+0.301 [+0.266, +0.336] ***
Full context (reference)+0.249 [+0.223, +0.273] ***+0.229 [+0.201, +0.258] ***
RAG+0.207 [+0.180, +0.234] ***+0.202 [+0.174, +0.232] ***
Rewrite+0.224 [+0.194, +0.256] ***+0.229 [+0.204, +0.254] ***
Mem0+0.195 [+0.171, +0.220] ***+0.188 [+0.167, +0.212] ***
A-MEM+0.207 [+0.182, +0.234] ***+0.233 [+0.204, +0.263] ***
MemListener+0.131 [+0.107, +0.153] ***+0.160 [+0.134, +0.187] ***
HabitusMem+0.242 [+0.212, +0.271] ***+0.267 [+0.236, +0.298] ***

On PersonaMem-v2, HabitusMem's gain lies above the confidence interval of every compact baseline under both agents, whereas with Qwen RAG does not improve on the base LLM and MemListener falls below it. On VitaBench 2.0, HabitusMem has the largest gain among compact memories under both agents.

Weakness 8: Source code

Thank you for the comment. The code is included in the supplementary material.

We hope that our responses have addressed your concerns, and we would be grateful if you would consider raising your score.

AuthorsOpenReview ↗

Followup on Rebuttal

Dear Reviewer;

Dear Reviewer; Thank you again for taking the time to review our responses. As the rebuttal period will be ending soon, we would greatly appreciate it if you could kindly revisit our responses and let us know whether any concerns or questions remain unaddressed.

We would be more than happy to provide further clarification or additional explanations if anything is still unclear. If our responses and the additional experiments have adequately addressed your concerns, we would also sincerely appreciate your reconsidering the overall score in light of the new evidence. Thank you again for your time and consideration.

Thanks

Reviewer_XU3sOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The user is represented as a latent preference state, and each interaction is a bayesian update on this state. The paper proposes HabitusMem, which is a writer that generates a belief about the user at its current state, in the form of text records that contain a claim with its evidence, scope, and status. The writer is trained with belief inference, belief update, and decision projection tasks. The resulting system outperforms existing memory systems on two long-horizon benchmarks.

Critical strengths

The paper studies an interesting problem of evolving user preferences, and identify issues in current semantic based memory systems for personalization. It provides some analysis of baseline methods and identify failure modes.

Critical weaknesses

I am not convinced by the paper’s main motivational argument, which states that existing memory systems store information of what the user said or did only, so they do not convey strength, context, staleness, and transient nature of claims (L40-51). It is possible to have a preference-aware summarizer or user representation without the Bayesian backbone to store all this information. The paper assumes personalization systems to only condition on exact historical conversation, which overlooks prior works on user modeling (e.g., https://arxiv.org/abs/2602.15012, https://arxiv.org/abs/2510.07925)

The paper’s main claim of novelty is the Bayesian belief memory, but the proposed model actually does not perform any real Bayesian updates. The bayesian framing is just used to frame the synthetic data generation process that produces synthetic data to do SFT on the rewriter. One of the challenges raised by the intro on transient preferences cannot be expressed by the proposed bayesian setup, which then results in the failed “request to forget” type of tasks.

The paper’s claim about HabitusMem tracking evolving latent preferences is not directly tested by the benchmarks used in the paper. PersonaMem and VitaBench mostly test stated updates (either explicit or implicit), but do not cover truly latent preferences that are results of life events. Benchmarks that target this setting, such as HorizonBench (https://arxiv.org/abs/2604.17283) and ImplexConv (https://arxiv.org/abs/2503.07018) should be used to directly test the rewriter’s ability to correct belief revision.

The section comparing prompting and training is extremely unclear. How exactly is the rewriter prompted? Or is the untrained rewriter directly conditioned on the request and compared to the trained rewriter? The claim is way stronger than the evidence provided in the paper.

Questions for authors

None

6 comments
AuthorsOpenReview ↗

Response to Reviewer XU3s (1/2)

Weakness 1: Motivation and prior work on user modeling

We disagree with two statements in this comment and add new evidence on the remaining question.

The paper does not claim that a Bayesian backbone is required to store these properties. Section 3.1 states that HabitusMem "neither estimates numerical posteriors nor optimizes Equation 2 through agent rollouts." HabitusMem is itself a writer that states evidence, scope, and status without computing a posterior. Our motivation concerns what existing systems convey, and the possibility of building such a summarizer does not show that existing systems convey these properties.

The paper does not assume that personalization conditions only on the exact conversation. The passage the reviewer cites (L40–51) discusses systems that "extract facts, maintain notes or summaries, or retrieve and compress past passages", and Section 2 of the submission discussed profile modeling, which builds user representations from persona descriptions. The evaluated conditions also include extracted-fact memory (Mem0), note-based memory (A-MEM), an LLM-rewritten summary (Rewrite), and, on VitaBench 2.0, a privileged ground-truth profile.

The two cited works do not provide these properties. CAPE (arXiv:2602.15012) addresses cold-start elicitation for users with no interaction history, states that it does not address longitudinal learning across sessions, and is itself built on online Bayesian inference. arXiv:2510.07925 keeps timestamped memories and an LLM-updated JSON profile of key facts. As described, neither records how strongly the history supports a claim, where the claim applies, or whether it was withdrawn. Among prior memory systems, Zep keeps validity intervals on facts, which addresses staleness but not evidential support or applicability. The revised Section 2 cites both works.

We evaluate both cited works as baselines.

Memory (PersonaMem-v2)Qwen Acc.DeepSeek Acc.
Base LLM (no memory)0.4460.266
Mem00.4700.412
A-MEM0.4890.408
Rewrite0.4840.438
Persistent profile (arXiv:2510.07925)0.4000.386
CAPE (arXiv:2602.15012)0.3010.193
HabitusMem0.5370.484

A profile of key facts that an LLM updates over the history trails HabitusMem by 13.7 and 9.8 points. Maintaining a user representation is therefore not sufficient by itself: the profile states facts about the user but not how well each is supported, whether it still holds, or where it applies. CAPE falls below the memory-free base LLM under both agents. It is designed to elicit preferences from a new user with no history, so its prior over structured criteria does not capture the specific preferences that a long history reveals.

Weakness 2: The Bayesian update

We agree that HabitusMem performs no numerical Bayesian update, as Section 3.1 states, and we have revised the title and claims accordingly (General Response, Part 2). We disagree with the other two statements in this comment.

The formulation does more than frame data generation. It specifies that memory should depend on the current decision and that each claim should carry its evidence, scope, and status. Tables G1 and G2 of the General Response test both choices with the training data held fixed.

Transient preferences and requests to forget are expressible in the formulation. The state ztz_t comprises durable dispositions, current circumstances, and explicit constraints (Section 3.1).

  • An occasion-only preference is a claim whose scope is that occasion. Section 3.2 gives an oat-milk order scoped to "this delivery only", which keeps its exact quantity without steering later requests.
  • A request to forget is an explicit constraint: it changes the loss of using a piece of information without making that information false. A probability of truth alone cannot express this, which is why the status field records the retraction separately (Section 3.1, App. A).

The forgetting errors arise in the agent's use of the records, not in what the formulation can express. The controlled probes (App. D.1, Table 6) change one element of the history and keep the question and options fixed. The full writer records the conditional scope and the retraction in all ten memories of each cell, yet the agents' answers differ:

ProbeFull writer, QwenFull writer, DeepSeekPlain-text writer, Qwen
Conditional scope, solo question10/102/107/10
Superseded preference9/109/108/10
Retracted preference1/102/100/10

The retraction is present in every memory, yet both agents usually ignore it. An explicit status therefore makes the constraint available but does not ensure that a frozen agent obeys it.

AuthorsOpenReview ↗

Response to Reviewer XU3s (2/2)

Weakness 3: Benchmarks for latent preferences

We disagree that the two benchmarks test only stated preferences. In VitaBench 2.0 the agent must complete delivery, in-store, and travel tasks according to preferences that are implicit in past dialogues and behavior logs, so the writer has to infer a preference that the user never states as an instruction (Section 4.1). PersonaMem-v2 includes preferences that change over the history and requests to forget.

We evaluated HabitusMem and the baselines on HorizonBench.

MethodDeepSeekQwen
Full context (reference)0.23270.2332
RAG0.19600.1829
Rewrite0.18820.2028
Mem00.22650.2632
A-MEM0.20350.2345
HabitusMem0.25490.2523

HabitusMem has the highest score with DeepSeek and exceeds full context under both agents. MemListener runs much more slowly on HorizonBench's long histories and its evaluation is still in progress.

Weakness 4: Prompting versus training

The revision states the protocol in Section 4.3 and App. B.3. Every lineage condition uses the same inference interface as HabitusMem: the belief instruction, the selected history, and, where present, the request suffix.

  • Base, static: the unadapted Qwen3.5-9B writer with the belief instruction and no request.
  • Base + request: the same unadapted writer with the current request in the suffix. The untrained writer is therefore conditioned on the request exactly as the trained writer is.
  • Base + unrelated: the same suffix with the fixed question "Recommend a good book on European history for a beginner" for every query.
  • Trained, static: the cold-start writer trained on teacher-written memories, evaluated without a request.
  • Parent writer: trained on the request-conditioned direct-writing examples; it initializes HabitusMem.

The submission reported this comparison only for DeepSeek on PersonaMem-v2. The revision reports both agents on both benchmarks (Tables 3 and 5):

WriterPM Qwen Acc.PM DeepSeek Acc.VB Qwen Avg@4VB DeepSeek Avg@4
Base, static0.3830.3290.25100.2700
Base + request0.3930.3450.34100.2897
Base + unrelated0.4010.3470.19400.2100
Trained, static0.4060.3350.26400.2480
Parent writer0.5050.4570.30120.3243
HabitusMem0.5370.4840.32910.3573

On PersonaMem-v2 the evidence supports the claim under both agents. Giving the unadapted writer the current request raises accuracy by no more than giving it an unrelated question, so prompting yields no benefit specific to the request, whereas the Parent writer and HabitusMem are 12 to 16 points above Base, static.

We hope that our responses have addressed your concerns, and we would be grateful if you would consider raising your score.

AuthorsOpenReview ↗

Followup on Rebuttal

Dear Reviewer;

Dear Reviewer; Thank you again for taking the time to review our responses. As the rebuttal period will be ending soon, we would greatly appreciate it if you could kindly revisit our responses and let us know whether any concerns or questions remain unaddressed.

We would be more than happy to provide further clarification or additional explanations if anything is still unclear. If our responses and the additional experiments have adequately addressed your concerns, we would also sincerely appreciate your reconsidering the overall score in light of the new evidence. Thank you again for your time and consideration.

Thanks

Comments on the paper

2 comments
AuthorsOpenReview ↗

## General Response (1/2): Summary of revisions

We thank the reviewers for their detailed comments and have uploaded a revised PDF. The main changes to the paper are listed below, followed by the new experiments reported in this discussion. Part 2 addresses the shared concern about the Bayesian formulation. Section, table, and appendix numbers refer to the revised PDF.

ChangeLocationReviewers
Removed "Bayesian" from the title, abstract, and contributions. Equations 1–2 are presented as the design rationale that specifies what the memory should state, and the text says explicitly that no posterior is estimated and that the writer keeps no persistent stateTitle, abstract, Sec. 1, Sec. 3.1–3.2, App. AXU3s, WKGL, uD1c
Expanded the component analysis: two request interventions on the trained writer, leave-one-out retraining for each training task, a Direct-only control with a matched optimization budget, an end-to-end Plain-text control for the record fields, both agents for every condition, and lineage controls on VitaBench 2.0Sec. 4.3, Fig. 3, Tables 3–5, App. B.3XU3s, uD1c
Defined each writer in the training lineage precisely, including how the unadapted writer receives the requestSec. 4.3, App. B.3XU3s
Added significance tests: McNemar tests for PersonaMem-v2 accuracy, paired bootstrap over task clusters for VitaBench reward, and McNemar tests for any-success and all-successApp. B.3WKGL
Added inference sensitivity to the history input budget, the history selector, the output cap, and the instructionApp. C.2WKGL
Added training cost, writer input and output length, evaluation time, and memory length relative to full historyApp. C.3WKGL
Replaced the observational retraction analysis with controlled probes that change one element of the history and separate what the writer records from what the agent usesApp. D.1, Table 6XU3s, uD1c
Added two case studies, one with a retracted preference and one with a paid delivery orderApp. E
Added related work on user modeling with Bayesian preference priors (CAPE) and persistent user profiles (Westhäusser et al., 2025), and on memory construction trained with reinforcement learning (MemAgent, Mem-α)Sec. 2XU3s, WKGL
Polished the figures for readability and visual consistencyAll figures
Included the code in the supplementary materialSupplementary materialWKGL

In addition, we ran the following experiments for this discussion. Their results are reported in our replies to the individual reviewers.

ExperimentReply
Persistent-profile memory (arXiv:2510.07925) and CAPE (arXiv:2602.15012) as baselines on PersonaMem-v2XU3s, Weakness 1
Evaluation on HorizonBenchXU3s, Weakness 3
MemAgent and Mem-α as baselines on PersonaMem-v2 and VitaBench 2.0WKGL, Weakness 5
Significance tests of every memory system against the no-memory base LLMWKGL, Weakness 7
Audit of retraction records and paired memory edits on PersonaMem-v2 forgetting queriesuD1c, Weakness 3
AuthorsOpenReview ↗

General Response (2/2): The Bayesian formulation

All three reviewers note that HabitusMem does not compute a posterior or optimize the risk in Equation 2. This is correct, and Section 3.1 of the submission already stated that HabitusMem "neither estimates numerical posteriors nor optimizes Equation 2 through agent rollouts". The title nevertheless suggested otherwise. The revision removes "Bayesian" from the title, which now reads "HabitusMem: Belief Memory over Evolving User Preferences for LLM Agents", and from the abstract and contributions. Section 3.1 presents Equations 1–2 as the design rationale that specifies what the writer should produce, and Section 3.2 states that the writer keeps no persistent state between requests.

The formulation is not decorative. It determines three design choices, and each is tested by a comparison in which only that choice changes.

(1) Memory depends on the current decision. The loss in Equation 2 depends on the request. Under a fixed budget, conditioning on the request cannot make the best attainable risk worse (App. A.3). With the trained writer fixed, we change only the request passed to it (PersonaMem-v2, Table 4).

Table G1.

Request passed to the writerQwen Acc.Qwen Gen.DeepSeek Acc.DeepSeek Gen.
Current request (HabitusMem)0.53720.32570.48420.3824
None0.40040.13450.33940.1995
Fixed unrelated question0.40190.12470.33760.1859

(2) A record states the evidence, scope, and status of a claim. The belief separates how credible a claim is, where it applies, and whether it still holds. The Plain-text control repeats both training stages with the same rows and settings and removes only these fields (PersonaMem-v2, Table 5).

Table G2.

Record formatQwen Acc.Qwen Gen.DeepSeek Acc.DeepSeek Gen.
Plain text0.5240.2930.4650.343
Claim with evidence, scope, status (HabitusMem)0.5370.3260.4840.382

(3) The writer infers, revises, and selects beliefs. These operations define the four training tasks. At the same number of optimization steps, training on all four tasks rather than on direct writing alone improves tool use (VitaBench 2.0, Table 3).

Table G3.

Training tasks (5,096 steps each)Qwen Avg@4DeepSeek Avg@4
Direct writing only0.29440.3136
All four tasks (HabitusMem)0.32910.3573

We therefore no longer attribute the gains to Bayesian inference at run time. They come from request conditioning, the record fields, and the training tasks, which the formulation motivates.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.