acceptodds
Under review as a conference paper at ICLR 2027

DualMem: Bridging Semantic and Episodic Memory for Contextual-Aware Privacy in Large Language Models

Abstract

As large language models (LLMs) increasingly act as persistent assistants, they are expected to use personal information to help users accomplish tasks. However, the appropriateness of such information varies by situation. Contextual Integrity (CI) characterizes this distinction through the actors involved, the information shared, the action taken, and the surrounding context. An LLM must therefore reason over these relationships before deciding how personal information should be used. We propose DualMem, a privacy-aware memory architecture that supports this process through two complementary memory types: Public Semantic Memory and Private Episodic Memory. Public Semantic Memory stores reusable knowledge about contextual privacy norms, while Private Episodic Memory captures situation-specific information states and relationships. Both memories are represented as event-centric hypergraphs to preserve higher-order relationships among CI elements. Given a task, DualMem retrieves relevant norms and grounds them against private information items using the current information flow and knowledge states, enabling context-aware privacy reasoning. We evaluate DualMem on three privacy benchmarks spanning probing, task-oriented generation, and CI compliance. Results show that structured semantic and episodic memory improves privacy recognition and compliance while preserving task helpfulness. Our analysis further reveals a gap between recognizing privacy constraints and consistently applying them during task execution.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Contextual Integrity, Human memory, Privacy-aware Reasoning, Large Language Models

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_XR9QOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DualMem, a privacy-aware memory architecture that separates generalized privacy norms and situation-specific information into public semantic and private episodic memories, represented as event-centric hypergraphs. It retrieves and grounds contextual norms to individual information items to guide LLM disclosure decisions. Experiments across three privacy benchmarks and eight LLMs demonstrate reduced privacy violations, with ablation studies examining the contributions of different components.

Critical strengths
  1. Practical and Meaningful Research Problem. The paper addresses an important privacy challenge in persistent LLM assistants, with clear applications to email management, agent tool use, and personalized services. Applying Contextual Integrity (CI) to memory-based privacy reasoning is a valuable research direction.

  2. Well-Structured Framework Design. The separation of generalized social norms into Public Semantic Memory and case-specific information into Private Episodic Memory is conceptually reasonable. In particular, distinguishing situation-level norm activation from item-level grounding enables more fine-grained privacy reasoning rather than applying the same privacy constraints indiscriminately to all information.

  3. Broad Experimental Coverage. The paper evaluates eight LLMs of different scales and architectures across three privacy benchmarks. Table 1 provides a comprehensive comparison of privacy protection and task utility, demonstrating the effectiveness of the proposed framework across diverse settings.

Critical weaknesses
  1. Unclear System-Level Privacy Threat Model. The paper does not sufficiently specify the trust boundaries of its memory architecture. In particular, it remains unclear who can access Private Episodic Memory, whether external extraction models can process sensitive user information, and how malicious or manipulated inputs might affect subsequent privacy decisions. A clearer threat model would help establish the security assumptions and practical limitations of the framework.

  2. Insufficient Baselines to Isolate the Core Contributions. DualMem provides the evaluated LLM with additional contextual analysis, extracted knowledge states, and item-level normative evidence. The current baselines do not adequately distinguish the benefits of the proposed memory architecture from those of stronger preprocessing and enriched prompting. Comparisons with structured-record baselines and reasoning-based privacy approaches would strengthen the claimed contributions.

  3. Insufficient Justification for the Hypergraph Representation. Although the paper emphasizes higher-order relationships in contextual privacy reasoning, the proposed method primarily relies on structured attributes and embedding-based similarity retrieval, without explicitly exploiting hypergraph topology or relational reasoning. It remains unclear whether the hypergraph representation offers advantages over simpler structured representations.

  4. Oversimplified Normative Reasoning and Conflict Resolution. The method prioritizes restrictive judgments over permissive ones, even when the latter have higher semantic relevance. This may lead to overprotection or unnecessary refusal, which does not fully align with the principles of Contextual Integrity. Furthermore, Social-Chem-101 contains general social norms that may not directly correspond to information-disclosure policies. The distinction between semantic relevance and normative applicability requires further investigation.

  5. Limited Evaluation of Grounding Reliability. The paper reports candidate norm coverage but does not sufficiently evaluate the precision or applicability of the final norm selected after status compression. This is particularly important when multiple conflicting norms are retrieved. Moreover, grounding quality is assessed using an LLM judge without independent human validation, raising concerns about evaluation reliability.

  6. Limited Evaluation of Persistent Memory Scenarios. Although the paper is motivated by persistent LLM assistants, Private Episodic Memory is primarily constructed from static, case-specific descriptions. The evaluation does not examine memory updates, evolving information states, consent revocation, or changing access permissions across interactions. This limits the evidence supporting the framework's applicability to realistic persistent-memory systems.

Questions for authors

I appreciate the authors' effort to bring Contextual Integrity (CI) into practical LLM memory systems, and I find the proposed framework conceptually reasonable. However, as detailed in the Weaknesses section, I remain concerned that the work does not sufficiently explore the complexity of contextual privacy reasoning or the challenges of realistic privacy scenarios. In particular, the justification for the proposed architecture and its ability to handle nuanced privacy boundaries require stronger evidence. I would be happy to reconsider my rating if the authors can convincingly address these concerns.

8 comments

We sincerely thank the reviewer for the constructive and detailed feedback. For clarity, all revisions in the manuscript are highlighted in blue text. We appreciate the recognition of the importance of contextual privacy, the separation of public semantic and private episodic memories, and the value of item-level grounding. The reviewer also highlights the breadth of our evaluation across eight LLMs and three privacy benchmarks. Below, we address the weaknesses (W) regarding the threat model, baseline comparisons, hypergraph representation, grounding reliability, and normative reasoning and persistent memory evaluation.

W1. System-Level Privacy Threat Model.

We agree that the trust boundaries were not stated explicitly. Public and Private describe the sharing scope of each memory (shared across cases vs. constructed per case), not a security mechanism, and we do not claim isolation or access-control guarantees. In our implementation, an external LLM (Claude-Sonnet-5) processes case information to construct Private Episodic Memory, so the extractor is assumed to be trusted.

DualMem targets inference-time disclosure decisions, not storage security of the extraction process. Unauthorized memory access and malicious input manipulation are not addressed. We expanded the Limitations (Appendix G), noting that practical deployments need complementary safeguards such as access control and secure memory management.

W2. Insufficient Baselines to Isolate the Core Contributions.

We thank the reviewer for raising this important concern. To separate the benefits of the memory architecture from those of stronger preprocessing and enriched prompting, we conducted additional experiments and expanded the ablation study.

First, we introduced CI-CoT (Contextual Integrity Chain-of-Thought), a reasoning-based baseline in which the evaluated LLM analyzes the actors, information, action, and context of the information flow using only its parametric knowledge, without external memory. The table reports averages across eight LLMs.

Benchmark / MetricNo-PolicyPrivacy-EnhancedNorm-RAGCI-CoTDualMem
PrivacyLens LR ↓53.835.935.133.724.8
PrivacyLens LR$_h$ ↓58.442.440.439.028.1
PrivacyLens Help. ↑2.552.452.492.452.47
ConfAIde LR ↓38.631.936.335.927.1
ConfAIde LR$_h$ ↓35.431.334.634.725.4
ConfAIde Help. ↑2.152.392.402.522.63
CIMemories Viol. ↓42.226.132.831.216.1
CIMemories Comp. ↑34.228.431.129.229.4

DualMem achieves lower average leakage than CI-CoT on all three benchmarks, with comparable or higher utility relative to CI-CoT. This suggests that explicit CI reasoning alone does not provide the same benefit as grounding reusable norms in situation-specific information.

Second, we report the extractor sensitivity experiment from the original submission (Appendix B.3, Table 10), now included in the main paper (Section 4.2, RQ4, Table 3), evaluating DualMem across different Private Episodic Memory extractors on 200 PrivacyLens cases.

ExtractorEvaluated LLMLR ↓LR$_h$ ↓Help. ↑Probe ↑
Claude-Sonnet-5*Mistral-7B-Inst41.543.82.4785.0
Qwen-3-8B-Inst42.044.02.6675.5
Qwen-3-8B-InstMistral-7B-Inst42.044.32.4384.0
Qwen-3-8B-Inst41.043.82.6074.5
Llama-3.3-70B-InstMistral-7B-Inst42.044.52.4684.5
Qwen-3-8B-Inst42.544.72.6375.0

Claude-Sonnet-5 uses explicit thinking in Stage A, whereas the other extractors do not. Relative to Claude-Sonnet-5, leakage and probing metrics change by at most 1.0 pp, suggesting that DualMem does not strongly depend on a particular preprocessing model.

Finally, the ablation study (Section 4.2, RQ3) isolates the contributions of Public Memory, Private Memory, structured episodic representation, and item-level grounding. Together, these experiments provide complementary evidence for the roles of explicit CI reasoning, memory construction, and the proposed architecture.

W3. Insufficient Justification for the Hypergraph Representation.

We agree that DualMem does not exploit hypergraph topology through message passing or graph-based reasoning. Hyperedges serve as relational units that jointly encode the attributes of a privacy event. To test whether joint encoding helps, we replaced each hyperedge with independently embedded relation-value pairs, aggregated by Norm ID before Top-5 retrieval.

VariantMistral LR ↓Mistral LR$_h$ ↓Mistral Help. ↑Mistral Probe ↑Qwen LR ↓Qwen LR$_h$ ↓Qwen Help. ↑Qwen Probe ↑
Full (DualMem)41.543.82.4785.042.044.02.6675.5
w/o Joint Hyperedge Encoding50.552.82.4274.551.053.02.5260.5

Removing joint encoding increases LR by 9.0 pp for both models and lowers probing accuracy by 10.5 and 15.0 pp for Mistral-7B-Inst and Qwen-3-8B-Inst, respectively. Independent embedding and aggregation of relation-value pairs thus does not preserve the benefit of jointly encoding hyperedge attributes. We clarified the role of hyperedges in Sections 1-3 and added this ablation to Section 4.2 (RQ3).

W4. Oversimplified Normative Reasoning and Conflict Resolution.

We agree that the restrictive-priority rule is a conservative heuristic for selecting supporting evidence, not a principled mechanism for resolving contextual conflicts. It may cause overprotection when a less restrictive norm is more applicable. Social-Chem also provides general social norms rather than disclosure policies, so semantic similarity does not guarantee normative applicability. The selected status is contextual evidence, not a deterministic rule, and the final decision remains with the evaluated LLM. We have added these points to Appendix G.

W5. Limited Evaluation of Grounding Reliability.

We thank the reviewer for highlighting the need to evaluate the final selected norms beyond candidate coverage. We conducted an additional grounding quality evaluation using GPT-5.4-mini on 50 PrivacyLens cases covering 358 information items. We separately assessed the semantic relevance of the retrieved candidate norms and the normative applicability of the final selected supporting norms.

MetricValue
Number of evaluated items358
Avg. # of Candidate Norms per Item4.40
No-match rate3.07%
Item Recall@572.51%
Item Precision@582.54%
Supporting Norm Applicability89.6%

The results show that 96.93% of information items have at least one candidate norm, with an Item Recall@5 of 72.51% and Item Precision@5 of 82.54%. Among items with an assigned supporting norm, 89.6% of the selected norms are judged applicable to their corresponding information flows. These results indicate that DualMem provides broad candidate coverage while generally selecting normatively applicable supporting evidence.

We have added these results to Section 4.2, RQ5 and provided the evaluation protocol and metric definitions to Appendix B.2 and C.2. This evaluation relies on an LLM judge without independent human validation, which remains a limitation.

W6. Limited Evaluation of Persistent Memory Scenarios.

We thank the reviewer for highlighting this important limitation. We agree that persistent-memory systems should support evolving information, memory updates, consent revocation, and changing access permissions, and our evaluation does not cover these. DualMem focuses on inference-time contextual privacy reasoning rather than the upstream construction and continuous maintenance of personal memory. These are complementary but distinct problems.

Since DualMem represents knowledge states as explicit hyperedges, updates to personal information or access permissions could in principle be reflected by modifying the affected nodes, attributes, and relations, without reconstructing the entire memory. Temporal knowledge graphs provide mechanisms for representing facts with temporal validity, while incremental KG construction supports updating knowledge graphs as underlying data sources change [1, 2].

However, reliable update mechanisms, particularly for consent revocation and evolving access permissions, require additional policies for temporal validity and consistency maintenance. We leave these to future work and state in Appendix G that dynamic memory construction and maintenance are beyond the scope of this study.

[1] Temporal Knowledge Graph Completion: A Survey, Cai et al., IJCAI, 2023,

[2] Incremental Knowledge Graph Construction from Heterogeneous Data Sources, Assche et al., Semantic Web, 2026.

Thank you for the detailed and thoughtful responses. I appreciate the authors' substantial efforts and recognize the challenges of incorporating Contextual Integrity (CI) into practical privacy mechanisms.

I believe W1, W2, and W5 have been reasonably addressed. However, I still have some concerns regarding W3 and W4.

For W3, the ablation supports joint encoding but does not fully establish the advantages of hypergraph representations over simpler structured alternatives.

For W4, since CI is central to this work, I believe its core principles deserve deeper investigation. CI concerns the appropriateness of information flows rather than simply restricting disclosure. The current restrictive-priority heuristic raises concerns about whether the framework can adequately handle conflicting norms and contextually appropriate sharing.

Regarding W6, I view dynamic memory management as a meaningful but challenging extension rather than a requirement for this work.

Overall, the rebuttal has strengthened the paper, and I appreciate the value of this research direction. However, given the remaining concerns, I will maintain my current rating.

Reviewer_PmJaOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents DualMem, a dual-memory architecture for mitigating contextual privacy leakage in LLM agents. It consists of (1) a public semantic memory that stores abstract contextual privacy norms and (2) a private episodic memory that captures situation-specific information, states, and relationships. Both memories are represented as event-centric hypergraphs to preserve higher-order relationships among contextual integrity (CI) elements, such as actors and information.

During response or action generation, DualMem retrieves information from both memories and performs item-level grounding, associating elements of the current situation with corresponding components of abstract privacy norms. The top-k relevant norms are then supplied to the agent as additional context.

In the implementation, public memory norms are extracted from Social-Chem-101, while private episodic memory graphs are constructed from benchmark case descriptions using Claude Sonnet 5. Experiments span three benchmarks (PrivacyLens, CIMemories, and ConfAIde) and both open-weight and proprietary models, ranging from Mistral-7B and Qwen3-8B to Claude Sonnet 5 and GPT-5.

Results show improvements in privacy probing accuracy, primarily for smaller models, and reductions in privacy leakage during actual actions across all evaluated models compared with the No-Policy, Privacy-Enhanced, and Norm-RAG baselines. Ablation studies on Mistral-7B and Qwen3-8B examine component contributions, and the paper also demonstrates the privacy–helpfulness trade-off.

Critical strengths
  1. Reasonable and practical approach. Retrieving contextual privacy norms relevant to the current situation to augment an agent's privacy reasoning is a straightforward and sensible approach.

  2. Effective item-level grounding. Rather than simply retrieving general norms, DualMem structures the current situation and explicitly associates its elements with corresponding components of contextual privacy norms. This is a useful design choice, and the improvements over Norm-RAG provide evidence of its effectiveness.

  3. Potential utility for smaller models. The approach appears particularly promising for augmenting smaller models, which may be more suitable for on-device deployment.

Critical weaknesses
  1. Dependence on a stronger model for private memory construction. All private episodic memories are extracted using Claude Sonnet 5. This introduces a potential confounding factor, particularly for improvements observed in smaller models. These models benefit not only from additional privacy norms and contextual grounding but also from situation-specific representations generated by a substantially more capable model, potentially leveraging its reasoning abilities.

  2. Simplified private memory construction that bypasses a central practical challenge. Private memory construction relies on benchmark-provided case descriptions, including narrative episode text, structured user profile attributes, and other structured information. This substantially simplifies the complexity of real-world agentic settings. In practice, information about social relationships, information accessibility, and who has access to particular information types is distributed across multiple interactions and sources rather than contained in a self-contained episode.

    Identifying and constructing these situation-specific representations is arguably the more difficult and consequential challenge. Once the relevant situational information and privacy norms are clearly represented, grounding and prompting the model to attend to those norms becomes relatively straightforward. Consequently, the paper does not adequately address what may be the harder and more practically important part of contextual privacy reasoning.

  3. Limited strength of the baseline comparisons. Norm-RAG is a meaningful baseline for demonstrating the benefits of contextual grounding beyond conventional norm retrieval-augmented generation. However, it should not necessarily be regarded as the strongest existing privacy-preserving approach. In particular, the evaluation lacks a comparison with explicit CI-aware reasoning, in which an LLM directly identifies actors, recipients, information states, and applicable privacy norms using its parametric knowledge rather than an external memory.

Questions for authors
  1. Role of the stronger model: To what extent do the improvements, particularly for smaller models, depend on the use of Claude Sonnet 5 to construct private episodic memories? How much of the benefit comes from DualMem's architecture versus the reasoning capabilities incorporated into the extracted representations?

  2. Realistic private memory construction: How would DualMem construct situation-specific memory representations when relevant information about actors, relationships, and information access is distributed across multiple interactions and sources, rather than provided in a structured, self-contained benchmark episode?

  3. Comparison with explicit CI-aware reasoning: How would DualMem compare against a baseline that explicitly prompts the LLM to identify relevant CI elements (e.g., actors, recipients, and information states) and reason about applicable privacy norms using its parametric knowledge, without external norm retrieval?

2 comments

We sincerely thank the reviewer for the detailed and constructive feedback. For clarity, all revisions in the manuscript are highlighted in blue text. We appreciate the recognition of DualMem as a practical approach to contextual privacy reasoning, particularly its item-level grounding of situation-specific information in relevant privacy norms. Below, we address the weaknesses (W) and the questions (Q) regarding private memory construction, extractor dependence, and baseline comparisons.

W1&Q1. Dependence on a Stronger Model for Private Memory Construction.

We thank the reviewer for raising this important concern. To test dependence on Claude-Sonnet-5, we moved extractor sensitivity experiment from the original submission (Appendix B.3, Table 10): we reconstructed Private Episodic Memories for 200 PrivacyLens cases using Qwen-3-8B-Inst and Llama-3.3-70B-Inst and evaluated them with Mistral-7B-Inst and Qwen-3-8B-Inst to main paper (Section 4.2, RQ4).

ExtractorEvaluated LLMLR ↓LR$_h$ ↓Help. ↑Probe ↑
Claude-Sonnet-5*Mistral-7B-Inst41.543.82.4785.0
Qwen-3-8B-Inst42.044.02.6675.5
Qwen-3-8B-InstMistral-7B-Inst42.044.32.4384.0
Qwen-3-8B-Inst41.043.82.6074.5
Llama-3.3-70B-InstMistral-7B-Inst42.044.52.4684.5
Qwen-3-8B-Inst42.544.72.6375.0

Claude-Sonnet-5 uses explicit thinking in Stage A; the other extractors follow the same two-stage procedure without explicit thinking.

Relative to the Claude-Sonnet-5 extractor, leakage and probing metrics change by at most 1.0 pp. Notably, the Qwen extractor, which is the same size as one of the evaluated models and uses no explicit thinking, achieves comparable results. These results suggest that DualMem's performance is not specific to Claude-Sonnet-5, although the effects of extractor capability and explicit thinking are not fully isolated.

This experiment does not fully separate extractor capability from structured extraction, so we also compare against CI-CoT (see W3&Q3), where the evaluated LLM performs CI analysis directly and DualMem still achieves lower leakage.

W2&Q2. Realistic Private Memory Construction.

We agree that this is a difficult and important practical challenge. In real-world agentic settings, information about actors, relationships, and information access is often distributed across multiple interactions and sources rather than provided as a self-contained episode.

In our evaluation, Private Episodic Memory is constructed from benchmark-provided narrative episodes and structured user profiles. This controlled setting lets us study how situation-specific information and knowledge states, once available, can be represented and grounded against reusable norms. Our contribution therefore concerns the grounding step: how already available situation-specific information can be combined with contextual norms to support information-flow decisions. It does not address collecting and integrating information from fragmented and evolving sources, which we regard as a separate and important problem. The knowledge-state hyperedges are source-agnostic, so in principle they could be populated incrementally from dialogue logs, tool outputs, or profile data, but we have not evaluated this.

We have expanded the Limitations section (Appendix G) to state this reliance on benchmark-provided information and the need for future work on multi-source construction, including integrating distributed interactions and maintaining evolving relationships and access states.

W3&Q3. Comparison with Explicit CI-aware Reasoning.

We thank the reviewer for this valuable suggestion. We added CI-CoT (Contextual Integrity Chain-of-Thought), a baseline that explicitly prompts the evaluated LLM to identify CI elements (actors, information, actions, and context) and assess the appropriateness of the information flow using its parametric knowledge, without external memory. The table reports averages across the eight LLMs.

Benchmark / MetricNo-PolicyPrivacy-EnhancedNorm-RAGCI-CoTDualMem
PrivacyLens LR ↓53.835.935.133.724.8
PrivacyLens LR$_h$ ↓58.442.440.439.028.1
PrivacyLens Help. ↑2.552.452.492.452.47
ConfAIde LR ↓38.631.936.335.927.1
ConfAIde LR$_h$ ↓35.431.334.634.725.4
ConfAIde Help. ↑2.152.392.402.522.63
CIMemories Viol. ↓42.226.132.831.216.1
CIMemories Comp. ↑34.228.431.129.229.4

CI-CoT improves over No-Policy, but DualMem lowers leakage further, by 8.9 and 8.8 points on PrivacyLens and ConfAIde and by 15.1 points on CIMemories relative to CI-CoT, with helpfulness comparable to or higher than CI-CoT on all three benchmarks. This suggests that explicit CI reasoning from parametric knowledge is less effective on average than grounding reusable norms in situation-specific information at the item level. We added CI-CoT to the main evaluation (Section 4.2, RQ2) and its prompt to Appendix C.1.

Dear Reviewer PmJa,

Thank you again for your detailed review. In our rebuttal, we responded to each of your questions: the extractor-dependence question (now reported in Section 4.2, RQ4, where replacing Claude-Sonnet-5 with smaller extractors gives similar results), the scope of private memory construction (clarified in Appendix G), and the comparison with explicit CI-aware reasoning (the CI-CoT baseline in Section 4.2, RQ2).

We would appreciate any further feedback on whether these responses address your concerns, and we are glad to continue the discussion.

Reviewer_oLxcOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DualMem, a privacy-aware memory framework for LLMs. The framework includes Public Semantic Memory and Private Episodic Memory. Both memories are represented using hypergraphs. During inference, DualMem retrieves relevant norms from public memory, grounds them against private information items, and provides contextual privacy guidance to the LLM. Experiments with eight LLMs demonstrate reductions in privacy leakage and improvements in contextual privacy reasoning.

Critical strengths
  • Intuitive memory architecture: Separating reusable privacy rules from personal information is reasonable
  • Broad empirical evaluation: The paper evaluates various LLMs across several privacy benchmarks, including component ablations and analyses of residual privacy leakage.
Critical weaknesses
  • Limited evaluation of privacy–utility trade-offs: DualMem can reduce privacy leakage, but helpfulness and completeness do not consistently improve. It is hard to infer the pratical usage of this method since we are not sure whether the method really achieve Pareto optimal.
  • Lack of comparisons with other privacy-aware memory methods: The experiments consider three baseline configurations, which primarily examine different levels of privacy guidance rather than alternative memory architectures. The paper does not evaluate against existing privacy-aware memory methods or adapted alternatives with comparable capabilities.
Questions for authors
  • the evaluation conditions are vaguely explained (line 339-343), might need citations or more formal definitions to elaborate?
  • lack of citations? may places (e.g. open and close-sourced models) do not have citations
4 comments

We sincerely thank the reviewer for the thoughtful feedback. For clarity, all revisions in the manuscript are highlighted in blue text. We appreciate the recognition of our intuitive memory architecture, particularly its separation of reusable privacy norms from situation-specific personal information. We also note that our evaluation covers multiple LLMs and privacy benchmarks, including component ablations and residual leakage analysis. Below, we address the weaknesses (W) and the comment (C) regarding privacy-utility trade-offs, baseline comparisons, and evaluation clarity.

W1. Limited Evaluation of Privacy-Utility Trade-offs.

We thank the reviewer for raising this point. Privacy leakage and task utility can be competing objectives, as withholding sensitive information may also limit task completion. The relevant question is therefore how much utility is traded per unit of privacy gained.

As shown in the revised results (in Section 4.2, RQ2. Does normative grounding reduce privacy leakage without substantially compromising task utility?), compared with No-Policy, DualMem reduces leakage by 29.0 and 11.5 pp on PrivacyLens and ConfAIde, with helpfulness changes of −0.08-0.08 and +0.48+0.48 (0-3 scale), respectively. On CIMemories, it reduces violations by 26.1 pp at a cost of 4.8 pp of completeness. This corresponds to about 5.4 points of violation removed per point of completeness lost, compared with about 2.8 for Privacy-Enhanced.

Moreover, DualMem achieves on average leakage among all five conditions on every benchmark. On ConfAIde it also achieves the highest helpfulness, and on PrivacyLens and CIMemories it attains helpfulness/completeness comparable to the other privacy-aware conditions (Norm-RAG retains slightly higher utility at substantially higher leakage). We have added these results to Section 4.2, RQ2.

W2. Lack of Comparisons with Other Privacy-Aware Memory Methods.

We thank the reviewer for this comment. To the best of our knowledge, there are no publicly available memory-based methods designed for contextual privacy reasoning in this setting, so we could not compare against an existing method under equivalent conditions. Instead, we compare against the closest adapted alternatives. Norm-RAG (already in the paper) retrieves social norms without information item-specific grounding. In this revision we add CI-CoT, which prompts the evaluated LLM to explicitly analyze the actors, information, action, and context of the information flow and to reason about its appropriateness using only its parametric knowledge, with no external memory. This tests whether the benefit of DualMem can be obtained by explicit CI reasoning alone.

The table below reports averages over the eight LLMs.

Benchmark / MetricNo-PolicyPrivacy-EnhancedNorm-RAGCI-CoTDualMem
PrivacyLens LR ↓53.835.935.133.724.8
PrivacyLens LR$_h$ ↓58.442.440.439.028.1
PrivacyLens Help. ↑2.552.452.492.452.47
ConfAIde LR ↓38.631.936.335.927.1
ConfAIde LR$_h$ ↓35.431.334.634.725.4
ConfAIde Help. ↑2.152.392.402.522.63
CIMemories Viol. ↓42.226.132.831.216.1
CIMemories Comp. ↑34.228.431.129.229.4

CI-CoT improves over No-Policy, but DualMem lowers leakage further, by 8.9 and 8.8 points on PrivacyLens and ConfAIde and by 15.1 points on CIMemories relative to CI-CoT. On average, explicit CI reasoning from parametric knowledge is therefore less effective than grounding reusable norms in episode-specific information at the item level. DualMem's helpfulness is comparable to or higher than CI-CoT's on all three benchmarks.

We note that CI-CoT is not an alternative memory architecture. It serves to isolate the construction of external memory and item-level grounding beyond explicit CI reasoning. We have added this baseline to Table 1 and Section 4.2 RQ2, and its prompt settings to Appendix C.1.

C1. Evaluation Conditions and Missing Citations.

We thank the reviewer for pointing this out. We have rewritten the Evaluation Conditions paragraph in Section 4.1 to define each of the five conditions by the guidance it adds to the original one: (i) No-Policy (task only), (ii) Privacy-Enhanced (an explicit privacy instruction), (iii) Norm-RAG (the instruction plus retrieved Social-Chem rules-of-thumb), (iv) CI-CoT (explicit reasoning over actors, information, action, and context using parametric knowledge, without external memory), and (v) DualMem (norm grounding over public and private memories at the item level). Exact prompt templates are given in Appendix C.1. We also reviewed the manuscript and added the missing references, in particular for the evaluated open- and closed-weight models.

Dear Reviewer oLxc,

Thank you again for your review. In our rebuttal, we addressed your two main concerns: (1) we added a privacy–utility analysis in Section 4.2 (RQ2), and (2) we added CI-CoT, an explicit CI-reasoning baseline, to the main evaluation. We also expanded the Evaluation Conditions in Section 4.1 and added the missing citations.

We would be grateful if you could take a look at our response and let us know whether it addresses your concerns, or if any points remain unclear. We are happy to discuss further.

Thank you for the rebuttal. Below are some comments:

  • I think in general this trade-off perspective is not discussed much in the paper. And in fact it is not achieving pareto optimal even just comparing against 2-3 baselines. I don't think this the proposed method has really improve significantly.
  • Newly added CI-CoT does not even have a memory structure. The additional addition adds zero to none justification of the proposed method's effectiveness. If the author is just trying to separate the role of reasoning and memory, DualMem's internal ablation should do the job.
  • Still, strong retrieval/memory baselines are missing. While I know this might take days of experiments and might be impossible to add during rebuttal, I feel this is a very important class of method that need to be included in the paper's scope.

Based on these limitations, I will maintain my score.

Thank you for your continued feedback. We would like to clarify our evaluation design on the three points you raised.

1. Privacy–utility trade-off. We agree that this perspective deserves more discussion in the paper, and we do not claim global Pareto optimality. Norm-RAG retains slightly higher utility on PrivacyLens and CIMemories, but at substantially higher leakage. We would also like to clarify the nature of this trade-off. On CIMemories, completeness measures the share of task-necessary information disclosed, while violation measures the share of private information disclosed. Since both are evaluated over the same user profile, reducing violations may also reduce completeness when the model withholds information that could contribute to task completion. Thus, the informative comparison is how much utility each method gives up for the privacy it gains, rather than whether it dominates every baseline on both axes. To make this trade-off more explicit, we have added a leakage–utility plot covering all five conditions in Appendix B.4 (Privacy–Utility Trade-off Analysis).

2. Role of CI-CoT. We agree that CI-CoT is not a memory-architecture baseline. It tests whether explicitly prompting the model to reason about contextual integrity, with information states inferred by the LLM itself, can achieve similar privacy improvements without externally retrieved norms or item-level grounding. The internal ablations address a different question: the contribution of each component within DualMem. Therefore, these experiments are complementary rather than substitutes.

3. Comparison with memory architectures. We agree that comparisons with adapted general-purpose retrieval or memory systems would further strengthen the evaluation, and we regard this as a valuable direction for future work. To our knowledge, no existing memory architecture is specifically designed for this setting, and adapting one would require additional design choices that could make direct comparisons difficult to interpret. Our evaluation includes Norm-RAG, which retrieves norms from the same public norm corpus without DualMem's episodic context and item-level grounding, and CI-CoT, which relies on the LLM's own contextual reasoning without external normative memory. Together with the controlled ablations and comparisons against independently embedded relation–value pairs and conventional pairwise KG edges, these experiments examine the key design choices underlying DualMem, particularly joint event representation and item-level norm grounding.

We thank the reviewer again for highlighting these points.

Comments on the paper

1 comment

Summary of the Discussion Period (Submission 35162)

Dear Area Chair,

We thank the reviewers for their careful reading. They recognized the importance of contextual privacy for LLM assistants, the separation of reusable norms from case-specific information, the item-level grounding design, and the breadth of our evaluation (eight LLMs, three benchmarks). Below we summarize the main concerns, our responses, and the remaining limitations. All revisions are highlighted in blue in the revised manuscript.

1. Is the gain from the memory architecture, or from stronger preprocessing and prompting? (PmJa, XR9Q, oLxc)

  • New baseline (CI-CoT): the evaluated LLM itself analyzes the actors, information, and information flow and judges their appropriateness under contextual integrity, without external memory. DualMem leaks less sensitive information than CI-CoT on all three benchmarks (average over eight LLMs), with comparable or slightly higher utility. Norm-RAG, which retrieves the same norms without case-specific grounding, shows the same pattern.
  • Memory-construction model: on 200 PrivacyLens cases, we rebuilt the private memory with two other models (Qwen-3-8B-Inst, Llama-3.3-70B-Inst) instead of Claude-Sonnet-5. Leakage rate (the share of responses disclosing sensitive information) and probing accuracy changed by at most 1.0 pp, so the results were nearly unchanged.

2. Is the hypergraph representation justified? (XR9Q)

  • We replaced hyperedges (which keep actor, action, information, and context together as one unit) with independently embedded attribute pairs and with pairwise knowledge-graph edges. Both performed worse: leakage rate rose by 6 to 9 pp and probing accuracy dropped by 10 to 15 pp.
  • We clarified that the benefit comes from keeping CI attributes together as one relational unit, rather than from hypergraph topology or message passing, and we scope this finding to the alternatives we tested.

3. Privacy–utility trade-off (oLxc)

  • DualMem achieves the largest privacy gain on all three benchmarks, while maintaining competitive utility. Although Norm-RAG achieves slightly higher utility on PrivacyLens and CIMemories, it incurs substantially higher leakage.
  • To assess both objectives jointly, we compare the aggregate ranks across the three benchmarks. DualMem achieves the best overall rank (10), combining the best privacy ranking (3) with a competitive utility rank (7), compared with Norm-RAG (18), CI-CoT (18), Privacy-Enhanced (21), and No-Policy (22).
  • We revised Section 4.2, RQ2 and added a benchmark-level analysis relative to No-Policy in Appendix B.4 to clarify how much utility is retained for the privacy gains achieved.

4. Grounding reliability and normative reasoning (XR9Q)

  • Beyond candidate retrieval, we evaluated the final selected norms: 89.6% were judged applicable to their information flows (Section 4.2, RQ5; Appendix B.2).
  • We acknowledge that the restrictive-priority rule (preferring the most restrictive norm) is a conservative design choice, not a principled method for resolving conflicting norms, and may overprotect. This is stated in Appendix G.

5. Scope and threat model (PmJa, XR9Q)

  • Private memory is built from benchmark-provided information. Multi-source construction, memory updates, consent revocation, and access permissions are outside our scope (Appendix G).
  • "Public" and "Private" denote sharing scope, not security isolation. An external LLM processes case information, and unauthorized access and input manipulation are not addressed.

State of the discussion

  • XR9Q agreed that W1, W2, and W5 were reasonably addressed, viewed dynamic memory management as an extension rather than a requirement, and raised the rating from 2 to 3 (weak acceptance) after our clarification of the design rationale.
  • oLxc maintained the score, citing that DualMem is not Pareto optimal against some baselines, the lack of memory structure in CI-CoT, and the absence of strong memory baselines. Regarding the first concern, Table 1 shows that no evaluated baseline outperforms DualMem on both leakage and utility on any benchmark (averaged over eight LLMs); baselines with slightly higher utility incur substantially higher leakage. Moreover, DualMem achieves the best aggregate rank across the six privacy and utility rankings. We agree that comparison with adapted general-purpose memory systems is valuable future work. To our knowledge, no existing method directly targets this setting, so Norm-RAG and CI-CoT are the closest baselines for each component.
  • PmJa has not responded yet. We answered all three questions: dependence on the extractor, realistic memory construction (scope), and the CI-aware reasoning baseline.

We thank the Area Chair and the reviewers for their time.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.