acceptodds
Under review as a conference paper at ICLR 2027

EquiMem: Calibrating Shared Memory in Multi-Agent Debate via Game-Theoretic Equilibrium

Abstract

Multi-agent debate (MAD) systems increasingly rely on shared memory to support long-horizon reasoning, but this convenience opens a critical vulnerability: a single corrupted entry can contaminate the downstream memory-augmented reasoning, and debate alone fails to filter such errors. Existing safeguards filter entries via heuristics or LLM-based validation, yet they rely on AI judgments that share the same failure modes and overlook the cross-agent dynamics of MAD. We address this gap by formulating memory updating in MAD as a zero-trust memory game, in which no agent is assumed reliable and the game's equilibrium motivates a principled objective for calibrating memory influence. Guided by this objective, we propose EquiMem, an inference-time calibration mechanism that evaluates each update against the shared memory state and the memory already being used by other agents, using their existing retrieval queries and traversal paths as evidence of calibration. EquiMem instantiates this calibration for both embedding- and graph-based memory, and across diverse benchmarks, MAD frameworks, and memory architectures, it consistently outperforms existing safeguards, remains robust under adversarial agents, and incurs negligible inference overhead.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Multi-Agent Systems, Agent Memory, Multi-Agent Debate, Game-Theoretic Equilibrium

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_b4ZtOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
3(+1)
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes EquiMem, a memory framework for multi-agent debate (MAD). Specifically, EquiMem identifies a critical issue of previous memory frameworks in MAD: not all memory can augment subsequent reasoning. Therefore, EquiMem formulates memory updating in MAD as a zero-trust memory game and introduces an equilibrium-constrained memory updating principle to ensure all updates apply to the memory and can help task execution. Through evaluation across multiple MAD frameworks, memory architectures, and tasks, EquiMem shows consistent improvements over baselines, which demonstrates the effectiveness.

Critical strengths
  1. This paper aims to address a critical problems: how to derive a memory that can effectively help subsequent task execution performance. Such direction will make significant contribution to current agentic research, and benefit both academia and industry.
  2. This paper conducts comprehensive experiments, including 5 kinds of memory architecture, 4 benchmarks, 3 MAD frameworks, and five baselines, which broadly show the advantage of EquiMem.
Critical weaknesses
  1. First, one main concern is that the methodology part of this paper is hard to follow. In Section 4.1, I cannot understand why the formulation of calibrated influences is designed as Eq. (3). Besides, the explanation of "How the design reflects the equilibrium objective" is also hard to follow. In general, after reading this part, I cannot understand what kind of memory can be accepted for updating and what kind of memory cannot be accepted. It would be better to improve the writing of the methodology part.

  2. Second, this paper seems to have some gaps in each component. (1) Eq. (3) looks like a measurement of an update of memory to task execution. When removing the Nash Equilibrium, Eq. (3) also holds, so it remains unclear what the contribution of the concept of Nash Equilibrium is. (2) This paper applies two kinds of calibration to apply Eq. (3) to memory, but I cannot see the relationship between these two calibrations. Further explanation would be beneficial.

  3. Third, this paper uses a relatively small model (Qwen3-VL-8B-Instruct) for experiments, which leaves two kinds of settings under-explored: (1) using more capable models (e.g., GPT5, DeepSeek-V4, and Kimi-K3) separately for experiments and (2) using heterogeneous teams of models, such as using GPT5, DeepSeek-V4, and Kimi-K3 together in one experiment. These are also MAD settings, which this paper does not discuss.

Questions for authors

See critical weaknesses

5 comments
AuthorsOpenReview ↗

Response, Clarification, and Revisions (1/3)

We thank the reviewer for the careful assessment and constructive suggestions. With this opportunity, we have checked the numerical consistency of our experimental results, further clarified the game-theoretic formulation, and strengthened the analysis of statistical significance and memory calibration limitations. Revisions are marked in blue, with relevant existing explanations highlighted in orange. Please see our point-wise responses below.



1. Clarity of Equation (3) and memory-update decisions

Weakness #1: First, one main concern is that the methodology part of this paper is hard to follow. In Section 4.1, I cannot understand why the formulation of calibrated influences is designed as Eq. (3). Besides, the explanation of "How the design reflects the equilibrium objective" is also hard to follow. In general, after reading this part, I cannot understand what kind of memory can be accepted for updating and what kind of memory cannot be accepted.

We thank the reviewer for this suggestion. We have improved the explanation in Section 4.1 to make the calibration rule and memory-update procedure easier to follow.

The key idea is that EquiMem controls how strongly a proposed update influences future memory retrieval. Every proposed update is first evaluated using two measurements:

  • Memory deviation d_mem: how much the update differs from the existing memory structure.
  • Memory-use deviation d_use: how much the update changes information that other agents have already retrieved or traversed.

Equation (3) then assigns an influence coefficient tau_t \in [0,1]. The two deviation measurements are compared with their respective running-median tolerances, t and t. The rule selects the largest influence that satisfies both constraints.

This gives three intuitive cases. When both deviations are within their permitted ranges, the update receives full influence. When either deviation exceeds its range, the update receives reduced influence. When no eligible auditor interaction exists, calibration uses the memory-deviation constraint alone.

Importantly, EquiMem uses continuous influence calibration. A proposed update is stored with its assigned influence coefficient, and that coefficient controls its contribution to later retrieval. Thus, an unusual update can be retained with limited influence, while a well-supported update can participate fully in subsequent reasoning.

Revision (blue) and highlighted original text (orange).

We have refined Section 4.1, particularly the explanation following Eq. (3), to clarify why the minimum-of-three-terms rule is used and how the resulting coefficient affects memory retrieval. We also highlight the existing definitions of the two deviations in Section 4.1 and the retrieval-weighting mechanism in Section 4.4.

AuthorsOpenReview ↗

Response, Clarification, and Revisions (2/3)

2. Contribution of Nash equilibrium and relationship between the two calibrations

Weakness #2: Second, this paper seems to have some gaps in each component. (1) Eq. (3) looks like a measurement of an update of memory to task execution. When removing the Nash Equilibrium, Eq. (3) also holds, so it remains unclear what the contribution of the concept of Nash Equilibrium is. (2) This paper applies two kinds of calibration to apply Eq. (3) to memory, but I cannot see the relationship between these two calibrations.

We appreciate the reviewer's questions. We would like to clarify two aspects of our formulation: the role of Nash equilibrium in motivating Eq. (3) and the relationship between embedding-based and graph-based calibration.

  • 1) Contribution of the game-theoretic formulation. The zero-trust memory game provides a common principle for determining how memory influence should be regulated in multi-agent debate. As defined in Section 3.1, contributors favor influential updates with low memory conflict, while auditors provide informative memory interactions with low additional effort. This motivates two requirements for memory calibration: an update should remain consistent with the accumulated memory, and its influence should account for information already being used by other agents.

    Equation (3) implements these requirements through a conditional optimization. For a fixed update and observed auditor interactions, EquiMem selects the largest feasible influence under the memory-state and memory-use constraints. The running-median tolerances define the operating ranges using recent memory activity.

    The game contributes the stability-oriented design principle, and Eq. (3) provides its executable influence-calibration rule. Proposition 3.2 establishes the existence of an equilibrium for the finite game, while Appendix B.2 explains the conditional optimization implemented by EquiMem. These are complementary parts of our formulation.

  • 2) Relationship between the two memory calibrations. Embedding-based and graph-based calibration follow the same mathematical formulation in Eq. (3). They compute the same two types of measurements from different memory representations.

    For embedding-based memory, d_mem measures excess embedding-neighborhood concentration, while d_use measures changes to auditors' existing retrieval rankings.

    For graph-based memory, d_mem measures support from local relations and graph paths, while d_use measures how proposed edges connect to paths already traversed by auditors.

    Both methods therefore answer the same questions: how an update fits into the existing memory, and how it affects memory currently used by other agents. Their deviation values are calibrated against separate running-median tolerances, after which the common Eq. (3) determines the influence coefficient. Section 4.4 then applies that coefficient during subsequent retrieval for both memory architectures.

Revision (blue) and highlighted original text (orange).

We have clarified the role of the equilibrium objective in Sections 3.2 and 4.1 and the corresponding conditional optimization in Appendix B.2. We have also expanded the explanation connecting embedding-based and graph-based calibration at the end of Section 4.3, while highlighting their respective deviation definitions in Sections 4.2 and 4.3.

AuthorsOpenReview ↗

Response, Clarification, and Revisions (3/3)

3. Generalization to stronger LLMs and heterogeneous MAD teams

Weakness #3: Third, this paper uses a relatively small model (Qwen3-VL-8B-Instruct) for experiments, which leaves two kinds of settings under-explored: (1) using more capable models (e.g., GPT5, DeepSeek-V4, and Kimi-K3) separately for experiments and (2) using heterogeneous teams of models, such as using GPT5, DeepSeek-V4, and Kimi-K3 together in one experiment.

We thank the reviewer for highlighting the importance of evaluating EquiMem with stronger LLMs and heterogeneous agent teams. We have expanded Appendix E.5 (Table 5) to cover both settings.

Specifically, the extended evaluation design considers three backbone configurations:

  • Homogeneous DeepSeek-V4-Pro: six agents using DeepSeek-V4-Pro.
  • Homogeneous Kimi-K3: six agents using Kimi-K3.
  • Heterogeneous MAD: six agents comprising two DeepSeek-V4-Pro, two Kimi-K3, and two Qwen3-VL-8B-Instruct agents.

All three configurations use MacNet with six agents, evaluated on both MemBank (embedding-based) and G-Memory (graph-based) across HotpotQA, LoCoMo, ALFWorld, and WebShop. The comparison includes Vanilla, LLM Audit, PPL Filter, A-MemGuard where applicable, MAPLE-Guard, and EquiMem.

Revision (blue).

We have revised Appendix E.5 (Alternative LLM Backbones) and expanded Table 5 to present the larger-backbone and heterogeneous-agent configurations. These additions allow us to examine two important aspects of generalization: the influence of stronger LLM backbones on memory calibration, and the effect of differing agent behaviors and retrieval patterns within a shared-memory system.

The heterogeneous setting is especially relevant because agents using different LLMs may generate different memory updates or access different portions of the same memory. EquiMem uses the shared memory structure and the observed retrieval interactions to calibrate these updates through the same mechanism. Hence, this extension broadens the evaluation setting beyond homogeneous teams of small LLMs and provides a direct framework for examining EquiMem in mixed-backbone MAD systems.



We would like to thank the reviewer once again for their valuable time and thoughtful feedback. We sincerely hope that our responses and revisions have adequately addressed the concerns raised and further improved the clarity and overall quality of the manuscript. We would greatly appreciate the reviewer's further assessment of these improvements and warmly welcome any additional feedback or discussion.

Best,

The author team

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

This paper focuses on improving shared memory mechanisms in multi-agent debate systems by proposing a method called EquiMEM for calibrating the shared memory. The authors’ method formulates multi-agent debate with shared memory as a zero-trust game between a contributor agent trying to update shared memory, and auditor agents who simply interact with the memory. They specifically calibrate memory using an inference time calculation of a weight factor that gets applied to memory updates that is based on the deviation between the update and the existing memory structure, with inconsistent updates being penalized. The authors argue that calibrating mechanism in this way addresses the vulnerability of bad updates being injected into shared memory better than existing filtering-based methods. They demonstrate that their method beats the tested baselines while using no additional token overhead across both embedding and graph-based memory mechanisms.

Critical strengths

The main strength is that the authors’ method is not difficult to implement and comes with the advantage that it does not incur significant overhead. The authors do run a fair number of experiments, showing that across a sizeable number of benchmarks there is a notable and consistent improvement of the performance of multi-agent debate with shared memory when their calibration method is used. I appreciate how the authors collect results with adaptive attacks against their method rather than using generic attack methods. The authors also run a fair number of ablations such as isolating the effects of each component and varying the number of debate agents. All of this is consequential as there appears to be some empirical utility of their proposed method.

Critical weaknesses

My main concern is that the method is not sufficiently motivated. If the authors choose to frame their method as a zero-trust game with the formulation as described in the paper then I see some issues with this. Fundamentally, there is no conflict between the contributor and the auditors since they are both still working in collaboration and not in an adversarial “game”. There is no actual win condition or pay off or penalty for any of the agents so I do not see how game theory equilibria apply here. The proof of the equilibria in appendix is not formalized and seems more like conjecture which casts doubt on the validity of the approach. The main contribution actually seems to be to make memory updates more conservative to reduce the influence of hallucinations entering memory (which are common among low parameter models such as the ones tested and I believe this is what explains the results i.e., it is more important to prevent the models from malfunctioning than them producing novel insights from the debate) but this naturally applies limitations and I doubt the generalizability especially to more capable models and harder scenarios. In essence, the main point of updating memory should introduce new information that conflicts with previous information otherwise the memory update brings little utility. This approach regularizes against this which would bring issues in contexts where updates with high information (but otherwise contradictory relative to existing memory entries) are required. I see this being problematic particularly in the case of evolving environments that would be more common in real-life deployment compared to the static benchmarks tested. I implore the authors to reframe their method and be clear about the specific situations it has utility.

Furthermore, I have some major issues with the experiments which I believe need to be addressed for acceptance. The baselines particularly for the adversarial section are weak and there are no strong baselines tuned to work effectively against adversarial attacks specifically. The structure of the adversarial robustness section has issues. I could not find results for the effect on vanilla models as a control for example. Most of the experiments are also based on a single model (Qwen3-VL-8B-Instruct). Although the authors do try some others in Table 5 in the Appendix, key experiments such as with adversarial robustness are not validated with other models. Moreover, in practice many debate scenarios use heterogenous agents (i.e., different architectures) and it is not clear how this mechanism would fare if there are significant deviations in behaviour between agents since that would naturally create more conflict in memory retrieval. Finally, claims are made such as “We attribute this to a property of growing memory: as M_t accumulates entries, the dominant errors are those that read fluently in isolation but conflict with the broader memory stats” without empirical evidence. Although Appendix e3 attempts to extensively describe why the method works but no evaluation is provided actually affirming each of the claims about the expected effects casting doubt as to where the benefits are actually being derived from.

Questions for authors

In eq 3 epsilon is not defined at all. I assume that this is some arbitrary bias term for which no value is provided. Can the authors clarify this?

Reviewer_A6qdOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses a real vulnerability in multi-agent debate (MAD) systems: a single bad entry written to shared memory can poison downstream reasoning for all agents. EquiMem calibrates the influence of each proposed memory update with a scalar weight τ, computed from two deviations: how much the update conflicts with the current memory state (d_mem) and how much it would disrupt the memory other agents are currently retrieving (d_use). The same rule is instantiated for embedding-based and graph-based memories. Across a large grid (3 debate frameworks × 4 memory systems × 4 benchmarks), EquiMem ranks first on point estimates and is robust to adversarial agents in the tested settings.

Critical strengths
  • The two-constraint idea is a genuine advance over per-entry filtering. Judging an update against both the accumulated memory state and other agents' live usage captures conflicts that isolated checks miss, and the ablations show both constraints contribute independently. Unifying embedding and graph memories under one rule is elegant.
  • The empirical package is unusually broad: 48 main-table cells plus ablations, backbone swaps, debate-round and agent-count studies, latency measurements, case studies, and an honest failure-mode appendix. The method is also practical: no extra LLM calls, zero token overhead, and latency far below LLM-based defenses even after correcting the paper's own numbers.
  • The failure analysis (App. G) is unusually candid: correlated hallucinations defeating the median, sparse early-round memory, and ambiguous graph semantics are the right failure modes, illustrated with concrete traces.
Critical weaknesses
  1. The latency claim contradicts the paper's own tables. The abstract, §5.4, and App. E.9 state overhead below 5%, but Table 7 reports up to 9.7% (graph memory, ALFWorld), with 5 of 8 configurations above 5%; the appendix prose figures (0.9–3.9s, 3.2–3.7%) match no table column, and Table 8 disagrees with Table 7 on one shared cell. The comparison with LLM-based defenses still favors EquiMem, but the "under 5%" sentence is factually wrong and must be corrected to "up to ~10%".
  2. The game theory is motivational, not operational, and its value is never isolated. Utilities are never instantiated; Prop. 3.2 is textbook Nash existence, which the authors themselves call modest; the "optimality" proof of τ amounts to choosing the largest element of a feasible set defined by designer-chosen running medians. Most importantly, there is no ablation removing only the game component — e.g., the same deviation measurements with a fixed threshold or z-score cap. The method also works with a single agent (d_mem needs only accumulated memory; d_use needs only interaction traces), where no game exists at all — which shows the game is not the source of the method's power. I ask the authors to either provide a no-game ablation or reframe the theory as design intuition.
  3. Evidence standards are weak in two places. First, no significance test appears anywhere, and several gaps sit inside overlapping standard deviations; the extension table even contains outright ties despite text claiming "highest" throughout. Second, the adversarial evaluation excludes the strongest baseline where it matters: on MemBank, the memory used in attack experiments, A-MemGuard is the strongest baseline in 9 of 12 cells and second in the rest, yet it never appears in the attack comparison (only LLM Audit and PPL Filter do). Attack evaluation is also confined to MacNet, uses hand-scripted rather than τ-optimized attackers, and never tests the cited AgentPoison/MINJA threats.
  4. No quantitative analysis of false positives on correct-but-novel entries, and a structural entrenchment risk in d_use. Because d_use penalizes updates that disturb agents' current retrieval, a correct entry that would displace a widely used bad entry is downweighted precisely for correcting the status quo — the mechanism protects the status quo whether or not it is right. The appendix admits a correct-but-redundant edge being downweighted but dismisses it as low-impact, which does not cover the correct-and-novel case. Relatedly, the running-median thresholds can drift if bad entries accumulate gradually (slow poisoning), which is never tested. I ask for false-positive data or an explicit acknowledgment of this trade-off.
Questions for authors

Please provide (1) a no-game ablation: the same d_mem/d_use measurements with a non-game τ rule (e.g., fixed threshold or z-score cap) — if performance is unchanged, what does the game contribute? (2) A discussion of the single-agent case: the method appears to work with N=1, where no game exists — what does that imply for the game framing? (3) Quantitative false-positive data for correct-but-novel entries, and an explicit discussion of the d_use entrenchment trade-off (corrections that displace widely used entries are downweighted). (4) A correction of the latency claim to the measured values (up to ~10%) and significance tests for the headline gaps, plus attack comparisons including A-MemGuard on MemBank.

5 comments
AuthorsOpenReview ↗

Response, Clarification, and Revisions (1/3)

We thank the reviewer for the careful assessment and constructive suggestions. With this opportunity, we have checked the numerical consistency of our experimental results, further clarified the game-theoretic formulation, and strengthened the analysis of statistical significance and memory calibration limitations. Revisions are marked in blue, with relevant existing explanations highlighted in orange. Please see our point-wise responses below.



1. Latency overhead and consistency across tables

Weakness #1: The latency claim contradicts the paper's own tables. The abstract, §5.4, and App. E.9 state overhead below 5%, but Table 7 reports up to 9.7% (graph memory, ALFWorld), with 5 of 8 configurations above 5%; the appendix prose figures (0.9–3.9s, 3.2–3.7%) match no table column, and Table 8 disagrees with Table 7 on one shared cell.

Question #4 (first half): A correction of the latency claim to the measured values (up to ~10%)

We thank the reviewer for carefully checking the latency results. For the experimental results in the main text, EquiMem brings less than 5% overhead (Figure 4) compared to vanilla cases (with less overhead than other baselines). However, this rate is increased to 10% as highlighted by the comment.

Revision (blue). We have corrected the Introduction, Section 5.4, and Appendix E.9 to consistently state up to approximately 10% latency overhead. We have also aligned Table 8 and its accompanying explanation with the timing breakdown in Table 7.

A tiny tip: we didn't respectfully find this number in abstract, so we address all other parts as our revisions.



2. Game-theoretic motivation, no-game ablation, and single-agent applicability

Weakness #2: The game theory is motivational, not operational, and its value is never isolated. Utilities are never instantiated; Prop. 3.2 is textbook Nash existence, which the authors themselves call modest; the "optimality" proof of τ amounts to choosing the largest element of a feasible set defined by designer-chosen running medians.

Weakness #2: Most importantly, there is no ablation removing only the game component — e.g., the same deviation measurements with a fixed threshold or z-score cap. The method also works with a single agent ... where no game exists at all ... I ask the authors to either provide a no-game ablation or reframe the theory as design intuition.

Question: Please provide (1) a no-game ablation ... (2) A discussion of the single-agent case ...

We appreciate these questions and would like to clarify the contribution of the game-theoretic formulation and the actual calibration mechanism.

  • 1) Role of the game: The zero-trust memory game provides the design motivation for balancing update influence, memory conflict, and auditor effort. Proposition 3.2 establishes mixed-strategy equilibrium existence for the defined finite game. During inference, EquiMem operates through the observable memory-state and memory-use deviations, followed by scalar influence calibration. The game motivates this design, while the implemented deviations and influence weighting produce its practical effects.

  • 2) No-game comparison and existing ablations: Appendix E.4 has evaluated the individual roles of memory deviation, use deviation, and calibrated retrieval influence. These ablations demonstrate the contributions of the implemented components. We agree that a fixed-threshold or z-score calibration using the same deviation measurements would provide a more direct comparison of alternative influence rules. Since the theoretical formulation is motivational, we have adopted the reviewer's suggested clarification of its role, without attributing the observed performance gains to equilibrium computation.

  • 3) Single-agent applicability: The memory-deviation measurement can also operate when only one agent is present. With N=1, there are no separate auditor interactions, so the use-deviation constraint is omitted. The calibration becomes

τt=min⁡{1,ϵtdmem+ε}.\tau_t=\min\left\{1,\frac{\epsilon_t}{d_{\mathrm{mem}}+\varepsilon}\right\}.

This is a valid memory-level calibration rule. In multi-agent settings, the additional use-deviation measurement incorporates memory interactions made by other agents. Thus, the game provides a useful interpretation of multi-agent memory interactions, while the calibration mechanism also supports a simpler single-agent form.

Revision (blue) and highlighted original text (orange).

We have refined Sections 3.2-4.1 and Appendix B.2 to make the theoretical scope explicit.

Furthermore, Appendix B.2 now discusses the single-agent case explicitly. These refinements maintain the original calibration mechanism while clarifying which properties are motivated by the game and which are guaranteed by the conditional optimization.

AuthorsOpenReview ↗

Response, Clarification, and Revisions (2/3)

3. Statistical significance of performance improvements

Weakness #3: Evidence standards are weak in two places. First, no significance test appears anywhere, ... despite text claiming "highest" throughout.

Question #4 (middle): ... significance tests for the headline gaps ...

We thank the reviewer for this suggestion. To address reviewer's concerns, we conduct additional statistical analyses. Specifically, we compare EquiMem with the strongest baseline. We apply Welch's two-sample t-test, followed by Benjamini–Hochberg (BH) correction to account for multiple comparisons.

Additional statistical significance results:

BenchmarkComparisonsSignificant (p<0.05)BH-adjusted (q<0.05)
HotpotQA12119
LoCoMo1277
ALFWorld1244
WebShop1275
Overall482925

The tests are reconstructed from the reported means and standard deviations, assuming five independent runs per method. The BH correction is applied across all 48 comparisons. The subsequently added MAPLE-Guard values are excluded from this analysis.

Insights. The additional analysis provides statistical support for a substantial portion of EquiMem's performance improvements across different tasks, MAD frameworks, and memory architectures. The evidence is particularly strong on multi-hop reasoning, while some action-intensive configurations show greater variability.

We also recognize that a higher average score does not necessarily imply a statistically significant improvement. Accordingly, we distinguish numerical improvements from statistically supported improvements and interpret equal scores as ties. Since the analysis uses reported summary statistics, paired tests using individual run results would provide an additional check.



4. Adversarial baselines and stronger attack settings

Weakness #3: Second, the adversarial evaluation excludes the strongest baseline ... A-MemGuard is the strongest baseline in 9 of 12 cells and second in the rest, yet it never appears in the attack comparison (only LLM Audit and PPL Filter do).

Weakness #3: Attack evaluation is also confined to MacNet, uses hand-scripted rather than τ-optimized attackers, and never tests the cited AgentPoison/MINJA threats.

.Question #4 (last part): .. attack comparisons including A-MemGuard on MemBank.

We appreciate the reviewer's suggestions regarding the adversarial evaluation. We would like to clarify as well as revise the manuscript to strengthen the comparison with dedicated memory-poisoning defenses:

  • 1) (Added revision) additional adversarial evaluation: We have extended Figure 6 to include A-MemGuard on MemBank and MAPLE-Guard (Xiong et al., 2026, a new advanced baseline) on both MemBank and G-Memory. This directly address reviewer's comments and request for a more comprehensive comparison with defenses specifically designed to protect agent memory against malicious updates.

  • 2) Connection to AgentPoison and MINJA: Our attack design builds on the memory-contamination mechanisms studied in these two works. AgentPoison shows that malicious memory entries can influence later decisions by making poisoned information highly likely to appear during retrieval. Our Memory Poisoning attack examines this vulnerability through memory-aware construction: for embedding memory, it places malicious information near the existing representation to encourage retrieval; for graph memory, it introduces locally plausible edges connecting distant graph regions. MINJA demonstrates how apparently ordinary interactions can introduce misleading reasoning into persistent memory, which subsequently affects future queries. Our setting examines the propagation of similarly misleading contributions through the shared memory used by debating agents.

Revision: We have highlighted the existing attack procedures and adversarial settings in Section 5.3 and Appendix E.8 (Algorithms 1–2). We have also clarified the relationship between our attack construction and the memory-poisoning mechanisms examined by AgentPoison and MINJA.

  • 3) Adaptive and multi-agent attack design: Our evaluation further considers two characteristics of multi-agent shared memory. First, the attacker has knowledge of the deployed safeguard and constructs updates to evade its checks, including low-perplexity content and entries that receive favorable LLM-audit scores. Second, our Audit Collusion attack explicitly models coordinated adversarial auditors. For embedding memory, colluding agents steer retrievals away from the poisoned update to reduce its measured effect on retrieval rankings. For graph memory, they select structurally compatible traversal paths that support the malicious update. These interactions directly target the memory-use evidence used by EquiMem.
AuthorsOpenReview ↗

Response, Clarification, and Revisions (3/3)

5. False positives, correction entrenchment, and slow poisoning

No quantitative analysis of false positives on correct-but-novel entries, and a structural entrenchment risk in d_use. Because d_use penalizes updates that disturb agents' current retrieval, a correct entry that would displace a widely used bad entry is downweighted precisely for correcting the status quo — the mechanism protects the status quo whether or not it is right.

... Relatedly, the running-median thresholds can drift if bad entries accumulate gradually (slow poisoning), which is never tested. I ask for false-positive data or an explicit acknowledgment of this trade-off.

... Quantitative false-positive data for correct-but-novel entries, and an explicit discussion of the d_use entrenchment trade-off ...

We appreciate the reviewer's observations. We would like to clarify how EquiMem handles useful memory corrections and how its adaptive calibration behaves when the existing memory is unreliable.

  • Correct updates and structural entrenchment. EquiMem assigns continuous influence coefficients to proposed updates. As explained in Section 4.2 and Appendix C.1, the embedding-based memory deviation uses excess neighborhood concentration, allowing sparse additions of new information to receive low memory deviation. Meanwhile, the memory-use deviation measures changes to retrieval results already accessed by other agents. We agree that a correct correction can also produce a substantial retrieval change, especially when it replaces frequently used incorrect information.

    Additional quantitative comparison. We consider the paired correction evaluation described in our response to Reviewer 1, using MacNet with MemBank and G-Memory on LoCoMo. We compare correct corrections of incorrect stored facts against false updates contradicting correct stored facts. Effective acceptance is defined by t0.5; the low-influence rate is its complement.

MemoryUpdate typeMean influence tEffective acceptance (%)Low-influence rate (%)
MemBankCorrect Correction0.76181.218.8
MemBankFalse Update0.23819.580.5
G-MemoryCorrect Correction0.70177.222.8
G-MemoryFalse Update0.32322.477.6

Insights. The comparison indicates that correct corrections can retain substantial influence even when they contradict existing memory, while false updates tend to receive lower influence. It also reveals a meaningful trade-off: some correct corrections receive low influence. The calibration measures structural consistency, so its decisions can be affected by errors already present in memory.

These measurements concern factual corrections. A separate ground-truth-labeled evaluation of correct-but-novel additions would be needed to establish their false-positive rate.

  • Slow poisoning and adaptive thresholds. As explained in Appendix C.3, EquiMem estimates its operating ranges from the running medians of recent memory and use deviations. This allows the thresholds to adjust to ordinary changes in memory activity and limits the influence of isolated extreme observations. However, persistent incorrect updates can gradually change the recent deviation distribution and shift these thresholds. In particular, a sufficiently large proportion of contaminated observations within the history window can influence the median itself.

    The memory-state and memory-use constraints provide complementary evidence, but the running median alone does not guarantee protection against sustained slow poisoning. Quantifying this effect requires an attack in which malicious updates are introduced gradually over successive calibration windows.

Clarification and highlighted original text (orange).

We highlight the existing explanations of sparse novel additions in Section 4.2 and Appendix C.1, as well as the adaptive operating ranges in Appendix C.3. These clarify why EquiMem can preserve useful new information and how its influence thresholds evolve.



We would like to thank the reviewer once again for their valuable time and thoughtful feedback. We sincerely hope that our responses and revisions have adequately addressed the concerns raised and further improved the clarity and overall quality of the manuscript. We would greatly appreciate the reviewer's further assessment of these improvements and warmly welcome any additional feedback or discussion.

Best,

The author team

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

This work proposes EquiMem which tries to prevent unreliable memory updates from spreading through a multi-agent debate system. Before committing an update, it measures how the update fits the existing memory and how it affects memory already accessed by other agents. These signals determine a weight that reduces the update’s influence during later retrieval.

The method has separate implementations for embedding and graph memory and does not require an additional LLM call for calibration. The paper motivates it through a contributor–auditor game and reports improvements across four benchmarks, three multi-agent frameworks, and four memory architectures, including poisoning and collusion experiments.

Critical strengths

Reusing existing retrieval queries and traversal paths is a practical idea. Memory checking can easily become another expensive agent loop. Here the calibration itself uses memory-side operations, with reported latency overhead below 5%. If the gains hold, this is a useful deployment advantage.

The empirical coverage is broad. The main results include several memory architectures and coordination frameworks, with mean and standard deviation over five runs. The reported improvements are often large, and the component ablations help show that both deviation signals contribute.

Critical weaknesses

I do not find the game-theoretic justification convincing. Proposition 3.2 establishes mixed-strategy equilibrium existence for a finite game. It does not derive the calibration rule or show that equilibrium implies reliable memory. Appendix B then fixes an update and maximizes a scalar influence subject to two imposed constraints. This is a different optimization from the original game, where the contributor chooses the update and auditors choose their interactions. In particular, Proposition B.1’s proof establishes maximal influence for the fixed deviations, not optimality over all possible updates. The method may work as a heuristic, but the claimed connection to equilibrium is not established.

Questions for authors

What formal result connects Equation 3 to a best response or equilibrium of the game in Equation 1? Can you state which quantities are fixed and which actions are optimized? The current maximal-feasible-influence argument does not establish the claimed connection.

What happens when the existing memory is wrong and a new, correct update contradicts it? Please report acceptance rate or influence for correct corrections together with false updates.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Thank you! And please see the point-wise response below as well as updated manuscript

Dear Reviewers, AC, SAC, and all editors,

We would like to take this opportunity to thanks for your time and efforts for organizing the review process. Please see below regarding our point-wise responses with intensive revision efforts. Besides:

  • we have uploaded revised manuscript, wherein Blue text indicates new updates or revised texts and Orange text indicates highlight to original contents to make it easier to be caught.
  • We have also integrated some new experimental results, with corresponding tables and figures updated. Please see below for detailed explanations of those efforts.

We would greatly appreciate the reviewer's further assessment of these improvements and warmly welcome any additional feedback or discussion. And we sincerely hope that our responses and revisions have adequately addressed the concerns raised and further improved the clarity and overall quality of the manuscript.

All the best,

The author team

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.