acceptodds
Under review as a conference paper at ICLR 2027

MemPermit: Governing Writes to Multi-Agent Shared Memory via Risk-Constrained Sequential Verification

Abstract

Shared long-term memory lets language-model agents reuse information, but each write commits both content and an audience. Existing admission filters, read-time access control, and fixed verification committees govern these choices separately. We introduce MemPermit, a write-time controller that selects persistence and a non-expanding reader audience under one prespecified, asymmetric risk profile. Risk triggers mandate specialist coverage; afterward, verification continues only when its expected reduction in action risk exceeds cost. On GateMem with MIRIX, the complete controller lowers access-control violations over read-time governance by 5.0 percentage points with GPT-4o-mini and 4.5 with DeepSeek-V4-Pro while meeting prespecified utility and over-refusal criteria. With identical specialist evidence, its joint terminal rule has a -point lower ACV estimate than an audience-blind two-stage rule. On held-out query-only MINJA attacks, it lowers attack success over its content-only detector by 9.7 and 5.4 points. Both model arms meet the declared -point judge-accuracy criterion on LoCoMo. Sequential verification reduces cost per candidate by 27% relative to a fixed committee and by 9-10% relative to one-shot routing under matched risk constraints.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM Agents, Multi-Agent Systems, Memory Governance, Access Control, Memory Security

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_DDyiOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents MemPermit, a write-time governance framework for shared memory in multi-agent LLM systems. The key idea is to jointly decide whether a candidate memory should be persisted and which readers should be allowed to access it. MemPermit combines mandatory risk-triggered specialist verification with a cost-aware sequential verification mechanism based on the expected value of verification (EVoV). It further enforces a non-expanding visibility scope so that future policy changes cannot broaden access beyond the audience determined at write time. Evaluation on three datasets and two models shows reduced access-control violations and memory-injection attack success while keeping benign utility degradation.

Critical strengths
  • Shared memory creates security risks that are not fully addressed by either content filtering or read-time access control alone. Treating persistence and audience selection jointly is a sensible formulation.
  • The paper considers both improved security and over-refusal and benign memory utility.
  • Extensive experiments cover access-control violations, memory injection, benign long-term memory utility, and verification cost, and the results seem sound.
Critical weaknesses
  • The framework is quite complex relative to the demonstrated gains and there's no code provided for reproducubility. MemPermit contains lightweight risk modules, RuleGuard, five specialist roles, belief updates, EVoV routing, reassessment, a joint aggregator, quarantine handling, and manually specified asymmetric losses. While each component is motivated individually, it is less clear whether this full complexity is necessary in practice. And a simpler end-to-end or learned write-time policy would be an important comparison.
  • The asymmetric loss in Eq. 3 assigns substantially different costs to harm, conflict, unauthorized disclosure, etc. And there is no clear justification regarding why they should be set in this way.
  • MemPermit assumes trusted identity and policy services, while compromise of the policy service or storage administrator is explicitly out of scope. This is understandable for scoping the work, but these components are particularly important because the proposed guarantees ultimately depend on their correctness.
  • All five specialists within each experimental arm use the same underlying model. The authors acknowledge that their marginal belief updates do not model correlations among same-backbone specialists. In practice, specialists based on the same model may make highly correlated mistakes, potentially making sequential verification appear more informative than it actually is.
Questions for authors
  1. How sensitive are the results to the risk weights in Eq. 3? Would the advantage of the joint persistence–audience policy remain under substantially different but still reasonable security/utility preferences?
  2. How much of the full MemPermit pipeline is actually necessary? Could the authors compare against a simpler learned or LLM-based write-time controller that directly predicts {reject, audience} using the same available information and verification budget?
  3. How correlated are the errors of the five specialists? Since they use the same underlying model, it would be useful to quantify agreement/error correlation and examine whether using heterogeneous models changes the effectiveness of sequential verification.
  4. How well does the framework scale with the number of principals? The current formulation maintains reader-specific authorization risks and considers audience selection explicitly. It would be useful to discuss or experimentally evaluate settings with much larger and more dynamic reader populations.
1 comment

Thank you for suggesting a direct LLM controller. We evaluated Direct-LLM on GateMem with GPT-4o-mini. MemPermit Full lowered the access-control violation rate (ACV) from 19.2% to 11.2%, with higher utility and fewer refusals of legitimate requests.

1. How sensitive is the decision to the loss weights?

In Eq. (3), rejecting a benign memory costs 10, retaining harmful content costs 100, and retaining conflicting content costs 20. Let wdisw_{\mathrm{dis}} be the cost of unauthorized disclosure and qgoodq_{\mathrm{good}} the product of the marginal probabilities that each of five risks is absent: harm, source unreliability, conflict, low retention value, and task mismatch. At qgood=1q_{\mathrm{good}}=1, with an exclusion cost of 2, including an authorized reader is preferable to excluding them when disclosure risk is below 2/(wdis+2)2/(w_{\mathrm{dis}}+2). For disclosure weights of 25, 50 (the paper setting), and 100, this threshold is 7.41%, 3.85%, and 1.96%. For any positive weights and fixed risk estimates, Full minimizes estimated loss over all feasible actions, including Decoupled's chosen action. Table 2 measures the empirical trade-off at the prespecified weights.

2. How does Full compare with a direct controller?

For each candidate, Direct-LLM outputs either rejection or the set of readers who may access the stored memory. Full uses the same backbone and jointly selects admission and audience from risk estimates.

Table R1 (new). GateMem, GPT-4o-mini.

ControllerACV ↓Utility ↑Over-refusal ↓USD/1,000 candidates
Direct-LLM (new)19.2%47.1%22.4%0.49
MemPermit Full (Table 1; Fig. 4)11.2%48.6%19.3%0.45

At similar observed cost, Full lowered ACV by 8.0 percentage points (pp), raised utility by 1.5 pp, and lowered over-refusal by 3.1 pp. The runs differed in cost by USD 0.04 per 1,000 candidates; both were below USD 0.50.

Table 2 isolates the decision rule under identical candidates, observations, and final beliefs. Decoupled decides admission from content risk before choosing an audience. Letting audience risk inform admission lowered ACV by 1.23 pp (Full minus Decoupled; two-sided 95% paired episode-cluster bootstrap CI: −2.11 to −0.47 pp), with utility down 0.9 pp and over-refusal up 1.5 pp. Compared with Admission-only, which must admit for all authorized readers or reject, Full lowered ACV by 2.8 pp, raised utility by 1.1 pp, and lowered over-refusal by 2.6 pp. Narrowing the audience lets Full retain a memory that is appropriate for only some readers.

3. What if specialists sharing a backbone make correlated errors?

Specialists judge different risks. A safe source judgment concerns reliability; a safe audience judgment concerns disclosure to a reader. Eq. (6) updates only the reported coordinates. Measuring shared errors requires comparing each role with its own target on the same candidate. Correlated errors could make another call appear more valuable than it is.

A mixed-backbone experiment would change the team. Fig. 4 holds the team fixed and changes routing. Comparing actual cost, ACV, and utility does not require independent specialist errors. For GPT-4o-mini and DeepSeek-V4-Pro, respectively, sequential verification saved 27.4% and 27.1% versus Static Full (all five specialists) and 10.0% and 8.5% versus One-shot. Across these four routing contrasts (sequential minus baseline), the one-sided 95% upper bound on the ACV difference was at most +0.95 pp, and the one-sided 95% lower bound on the utility difference was at least −1.8 pp, within the +1/−3 pp margins.

4. How does the system handle more readers and changing permissions?

Given authorization and per-reader risks, Eq. (3) selects an audience in O(∣A∣)O(|A|) time for the requested reader set AA. One scan includes all authorized readers whose inclusion lowers loss. If all increments are nonnegative, the reader with the smallest increment gives the best nonempty audience; we compare its loss with rejection. An empty authorized subset yields rejection. Larger requests also incur per-reader authorization, risk-estimation, and storage costs.

At retrieval, Eq. (8) intersects the stored audience with current authorization. Revocation blocks subsequent reads; later grants do not expand the stored audience. Identity and policy services supply the authenticated reader and current permissions, and the backend enforces the intersection. Missing identity or unresolved policy results in denial.

Reviewer_PN8HOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper introduces the MEMPERMIT, which is a write-time controller that selects persistence and non-expanding reader audience. The paper tested it on different datasets including the MIRIX with GPT-4o-mini, GOT 3.5 and other LLMs. It finds that it lowers the ACV estimate. It also states that the sequential verification reduces the cost per candidate by 27% relative to a fixed Committee.

Critical strengths
  1. The security of the multi-agent memory system is an important research topic.
  2. The risk aware verification and access control is interesting.
Critical weaknesses
  1. The readability of this paper is poor. For example starting from the abstract, it is hard to understand lots of new words here “a write-time controller that selects persistence and a non-expanding reader audience under one prespecified, asymmetric risk profile” and “Risk triggers mandate specialist coverage; afterward, verification continues only when its expected reduction in action risk exceeds cost.”. So it is very hard to understand what this paper did.
  2. In the problem formulation, it seems like the paper customized the memory a lot rather than following the conventional multi-agent memory design. For example, it has the authenticated source, requested set of reader principals and versioned authorization policy. Such customized multi-agent system may benefit the design of the paper itself while lacking the general value to the audience of multi-agent.
  3. The paper mentioned that they have a complex deployment policy through the asymmetric expected loss, while it is hard to understand what the deployment policy is. Why we need such deployment policy. What the asymmetric expected loss is as the paper is working on the agent rather than training.
  4. The overall motivation of introducing such customized agent and the method is not justified clearly in the paper. Also, it is hard to understand the position of the paper given the limited existing literature review, e.g., multi-agent system, memory, security etc.
Questions for authors
  1. Could the authors improve readability by clarifying the technical terminology and providing a more intuitive explanation of the proposed method?
  2. How generalizable is the proposed approach to conventional multi-agent memory systems without the customized components and assumptions?
  3. Could the authors clarify the motivation, necessity, and practical role of the asymmetric expected loss function and deployment policy?
  4. Could the authors better justify the motivation and novelty of the proposed framework and its positioning relative to existing research?
2 comments

Thank you for pointing out how much terminology the abstract asks readers to absorb at once. The question behind MemPermit is simple: when a memory is useful to one agent but too sensitive to share with another, can the system keep it for the right readers?

1. What do the controller and its verification rule do? Consider a hospital assistant. A handover note may be useful to the physician on duty, while its clinical details are inappropriate for a scheduling assistant. Discarding the note loses useful information; placing it in a common pool risks disclosure. MemPermit offers a third action: keep the note for a specific set of readers. A later addition to the workspace does not grant access to this old note, and a physician whose access is revoked can no longer read it. This is what we mean by a non-expanding audience.

Verification serves that decision. A request involving multiple readers triggers a mandatory audience check. After required checks, the controller seeks more evidence only when its expected reduction in decision loss justifies the cost. Compared with calling all five specialists, this rule reduced cost by 27.4% and 27.1% on the two backbones while meeting the prespecified access-risk and utility criteria (Fig. 4).

2. What does a conventional memory system need to adopt it? The reusable component sits immediately before a memory is committed. It needs three pieces of information from the host: who submitted the record, whom the writer requested as readers, and whom the current policy permits. For example, an application with authenticated callers, workspace membership, and access-control lists can supply these through existing interfaces. It must also store the chosen audience and enforce it at retrieval. An anonymous shared store would first need identity and authorization interfaces to provide reader-specific protection.

Our implementation uses MIRIX as the memory backend. Extraction, storage, ranking, and answering are shared across conditions. GateMem, MINJA-QA, and LoCoMo are the evaluation datasets; GPT-4o-mini and DeepSeek-V4-Pro are the two backbones.

3. Why use an asymmetric loss when the agent is making a decision, not training? Eq. (3) scores actions at inference time. A risk estimate says how likely an error is; the deployment policy specifies its consequence. For example, we assign a cost of 50 to exposing a record to an inappropriate reader and 2 to withholding it from an appropriate one. These relative costs are fixed before testing.

For a record with qgood=1q_{\mathrm{good}}=1, consider two readers allowed by the current policy. At 1% disclosure risk, the reader-specific loss is 0.50 for inclusion and 1.98 for exclusion, so inclusion is preferable. At 10% risk, the losses are 5.00 and 1.80, so exclusion is preferable. The controller then compares the best scoped write with rejecting the record altogether. Thus the weights have an inspectable consequence for each write.

4. What does this add to memory management and access control? Section 2 discusses systems such as Mem0 and MIRIX that organize and retrieve memories, shared-memory systems that regulate who has access, and filters that assess harmful content. MemPermit's contribution is to let the risk of the best feasible audience feed back into whether the record should persist at all. A handover note can be accurate and useful while still being inappropriate for one of its requested readers.

Table 2 tests this distinction directly. With identical candidates, specialist evidence, and final risk estimates, the joint rule reduced access-control violations by 1.23 percentage points relative to deciding persistence first and audience second (paired 95% interval for the reduction: 0.47 to 2.11 points), with a 0.9-point decrease in utility. At the system level, adding write-time governance to read-time enforcement reduced violations from 16.2% to 11.2% on GPT-4o-mini and from 15.1% to 10.6% on DeepSeek-V4-Pro, meeting the prespecified utility and over-refusal criteria on both (Table 1). The first comparison isolates the decision rule; the second measures the benefit over an existing access-control layer.

Reviewer_PN8HOpenReview ↗

Thanks for the Response

Thanks for the author's response. Overall, I think the paper's writing and visualization should be greatly polished in order to meet the high-quality standard for this conference and I will maintain my rating.

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

The paper proposes MemPermit, a write-time controller for shared memory in multi-agent systems. It jointly selects persistence and a non-expanding reader audience under an asymmetric loss. Mandatory specialist checks are followed by optional verification when the expected reduction in decision risk justifies its cost. Retrieval intersects the stored audience with the current policy. Evaluations on GateMem, MINJA-QA, and LoCoMo report improvements in access control and resistance to query-only memory injection within predefined utility tolerances, together with lower verification costs on two model backbones.

Critical strengths
  1. The paper connects persistence, reader scope, and evidence acquisition at the shared-memory write boundary. The asymmetric loss and the retrieval intersection provide an explicit account of the intended governance policy.
  2. It has positive results with a structured evaluation: On GPT-4o-mini, Full reduces GateMem ACV from 16.2% to 11.2% relative to Read-time, while the reported utility and over-refusal bounds meet the predefined criteria. The second model shows the same direction of improvement. The evaluation uses episode-grouped outer/inner splits, paired cluster bootstrap, and Holm correction, and reports the costs 3) The results are convincing according to the authors. on theRelative to Decoupled, Full reduces ACV by 1.23 percentage points (95% paired interval [-2.11, -0.47]), with Utility decreasing by 0.9 and over-refusal increasing by 1.5 points. The authors appropriately interpret this as a change in the operating point, which helps identify the role of audience-risk feedback.
Critical weaknesses
  1. The paper describes the control flow and gives RuleGuard examples, but does not specify the complete trigger mapping, observation-bin thresholds, or the operational criteria for triggering reassessment and identifying unstable results. These definitions determine expert invocation and admission, rejection, or quarantine outcomes, and are needed to reproduce the evaluated policy.

  2. One-shot and cost accounting need clarification (Section 5 and Section 6.5). Candidates, observations, aggregation, and failure handling are matched across routing methods, but the selection rule used by One-shot is not described. The cost accounting for Fig. 4 also leaves it unclear whether initial signals, retries, and reassessment are included. These details are needed to interpret the reported savings over One-shot.

Questions for authors
  1. Please provide the complete RuleGuard configuration, observation cut points, reassessment criteria, and the construction of the binary targets used in Eq. 5. In Eq. 7, does each predictive branch update only the primary risk coordinate, or can it update several coordinates as in the actual specialist observation?

  2. What selection rule and training procedure are used for One-shot? Which components are included in Fig. 4 costs? Reporting average expert calls and retry/reassessment frequencies would help explain the source of the savings.

  3. For MINJA, please report the attack-instance and template-family counts and the denominators of the attack and benign outcomes. Please also list the hypotheses grouped for Holm correction and the confidence level of the one-sided non-inferiority bounds. These details would identify the independent evaluation units and help reproduce the statistical inference.

1 comment
AuthorsOpenReview ↗

Clarifications on Reproducibility, Costs, and Statistical Analysis

Thank you for these questions. All equation numbers refer to the submitted paper.

1. What triggers checks, reassessment, and belief updates?

Triggers are combined by union. Source: external provenance, missing/conflicting reliability evidence, or ps≥τsp_s\ge\tau_s. Audience: ∣A∣>1|A|>1, sensitive readers, A+≠AA^+\ne A, or any pu,r≥τup_{u,r}\ge\tau_u. Coherence: NLI-initialized pc≥τcp_c\ge\tau_c. Adversarial content: external provenance or ph≥τhp_h\ge\tau_h. Task state: pl≥τlp_l\ge\tau_l, ptm≥τtmp_{tm}\ge\tau_{tm}, or duplication similarity ≥τdup\ge\tau_{dup}.

After provenance/audience/policy validation, mandatory checks precede optional calls. Retry invalid responses once; quarantine after another failure. NormalStop requires mandatory coverage, valid beliefs, a feasible action, and γ=[R(d1,b)−R(d0,b)]/10>η\gamma=[R(d_1,b)-R(d_0,b)]/10>\eta, where d0,d1d_0,d_1 are the best and second-best feasible ordinary actions; γ=∞\gamma=\infty for one feasible action. When no optional specialist has positive net value or all are exhausted, failure of NormalStop triggers two-pass reassessment: locate evidence/conflicts, then verify against original material. Only the second-pass observation updates beliefs. Invalid or still-unstable results are quarantined.

For Eq. (5), yk=1y_k=1 denotes harm, source unreliability, contradiction without a valid update, low value, task mismatch, or unauthorized disclosure. Labels use task facts, trusted provenance, and versioned policies; semantic labels are independently annotated with disagreements adjudicated. Unresolved labels are excluded per coordinate. Outputs in [0,l)[0,l), [l,u)[l,u), and [u,1][u,1] map to safe, uncertain, and unsafe. Observation/label counts receive Laplace smoothing.

Four-fold grouped inner validation selects cut-point quantiles from 0.05, 0.10, ..., 0.95 by binary log loss. Triggers and η\eta use this grid, adding zero for η\eta; coordinate-wise search minimizes cost subject to ACV increase ≤1\le1 pp and Utility loss ≤3\le3 pp. Cut points/likelihoods are refitted on outer-training data and frozen before testing. MINJA is excluded from training/calibration/prompt selection.

Each Eq. (7) predictive branch updates only the primary coordinate using Eq. (6), then re-solves Eq. (4). Actual observations update all reported valid coordinates. Primary coordinates are ps,pc,php_s,p_c,p_h for source/coherence/adversarial content; the audience reader minimizing ∣Δr∣|\Delta_r| in A+A^+; and the larger of pl,ptmp_l,p_{tm} for task state. They are reselected after each call.

2. How does One-shot work, and what costs are included?

All routes share candidates, specialists, mandatory rules, observations, aggregation, failure handling, and billing denominators. After mandatory checks, One-shot computes remaining specialists' net values once and executes the fixed positive-value set. EVoV calls the current maximum and recomputes after each update. Both reuse calibration, cut points, likelihoods, and cost estimates. Static Full runs all five specialists.

Costs include paid initial signals, routing, specialists, retries, and reassessment per all gate candidates, including rejections/quarantines/failures. Local CPU/GPU time is separate. Figure 4 savings for GPT-4o-mini/DeepSeek-V4-Pro are 27.4%/27.1% versus Static Full and 10.0%/8.5% versus One-shot. The four comparisons have ACV upper bounds ≤0.95\le0.95 pp and Utility lower bounds ≥−1.8\ge-1.8 pp, within +1/-3 pp margins.

3. What are the MINJA counts and statistical units?

Each model uses 9 families, each with 10 injection attacks, 10 separate attack-evaluation queries, 10 benign QA questions, and 30 benign injection queries, repeated over 20 runs. ASR/AtkAdm/BenAcc denominators are 1,800 each; BenAdm is 5,400. AtkAdm counts all attack-origin gate candidates, one per injection query. Off/B2-only/Full share query sets and denominators, retaining rejections/quarantines. Table 3 ISR (75%/55%) measures upstream injection success, distinct from admission.

Each of 10,000 paired bootstrap resamples draws 9 families with replacement, retaining all queries/runs and condition pairing. GateMem/LoCoMo use episodes/conversations. One-sided confidence is 95% (5th/95th percentile lower/upper bounds).

Holm is applied separately by RQ, including both models; Δ\Delta denotes Full/EVoV minus baseline; pp denotes percentage points:

  • RQ1 (6): Full minus Read-time: H0:ΔACV≥0H_0:\Delta ACV\ge0, ΔUtility≤−3\Delta Utility\le-3 pp, ΔOR≥+3\Delta OR\ge+3 pp.
  • RQ2 (10): MINJA Full minus Off/B2-only: ΔASR≥0\Delta ASR\ge0 each; Full minus Off: ΔBenAcc≤−3\Delta BenAcc\le-3 pp and ΔBenAdm≤−3\Delta BenAdm\le-3 pp. LoCoMo Full minus Off: ΔJudge≤−3\Delta Judge\le-3 pp.
  • RQ3 (8): EVoV minus Static Full/One-shot: ΔACV≥+1\Delta ACV\ge+1 pp and ΔUtility≤−3\Delta Utility\le-3 pp each.

Each inequality defines a null hypothesis; the alternative reverses it strictly. One-sided p-values use bootstrap differences centered at the null boundary, then Holm step-down correction.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.