acceptodds
Under review as a conference paper at ICLR 2027

Remembering More, Risking More: Longitudinal Safety Risks in Memory-Equipped LLM Agents

Abstract

Persistent memory makes agent safety a longitudinal property: an ordinary request can become unsafe because of state accumulated much earlier. Measuring this effect is difficult because cumulative failure counts also change when the interaction stream itself changes. We introduce a trigger–probe protocol that evaluates fixed benign requests against persistent states reconstructed at increasing exposure levels, with matched memory-free counterfactuals for operational attribution. Across configurable memory architectures, three model backbones, and synthetic and real correspondence streams, memory-induced violation rates generally increase with accumulated experience. In persistent OpenClaw workspaces, we additionally distinguish retained state from later outcomes and find that sensitive values can continue spreading across workspace artifacts even after the set of retained values has stabilized. This decomposition exposes an observable intervention point for retrieval-mediated agents: the context selected from memory is available before generation. We package the longitudinal evaluation workflow as MemVax.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM agents, memory safety, privacy, long-horizon evaluation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_VDG5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies safety risks from persistent memory built through benign interaction histories. Its trigger–probe protocol tests fixed requests across exposure levels, using matched NullMemory responses and retrieved evidence to identify memory-induced violations. Experiments cover eight memory architectures, three backbones.

Critical strengths

(1) The longitudinal evaluation allows the reported safety trend to be read over time.

Fixed probes hold requests constant as memory changes, and matched NullMemory responses with retrieved evidence help attribute the observed effects to memory

(2) The study covers different memory settings and separates distinct failure stages.

Office-assistant evaluations span architectures and backbones. OpenClaw separates credential retention, activation, propagation across files, and unauthorized disclosure.

Critical weaknesses

(1) The novelty is limited, and the causes of the reported increase in memory-related risk remain insufficiently substantiated.

A. Benign histories do not necessarily establish a new risk mechanism. AgentPoison (Chen et al., 2024) and A-MemGuard (Wei et al., 2025) examine how stored information affects later behavior, while CIMemories (Mireshghallah et al., 2025) studies inappropriate disclosure from persistent memories without malicious instructions. Although risky information enters memory differently, these works share a central risk: stored content is retrieved and misused outside its appropriate context. The absence of attacker-crafted input distinguishes the setting, not necessarily the failure mechanism. The paper should identify which findings are unique to benign-history accumulation and cannot be explained by prior memory-security research.

B. The methods have limited technical novelty, and experiments cannot fully validate their value. The trigger–probe protocol combines fixed requests, checkpoints, and memory-free comparisons; Event-Structure prompts an existing LLM about requests and retrieved items. Its judgments may contain hallucinations or unsupported inferences, but the experiments do not establish that predicted risk reflects the proposed event mechanism. Nor does Table 1 demonstrate superiority: on Registrar, Event-Structure's F1 (0.573) trails CI Checklist (0.581), despite high recall. Thus the results do not establish a methodological advance or effective defense.

C. The increasing violation rate does not identify why risk increases. Longer histories may increase retrieval of risk-enabling records without increasing misuse conditional on retrieval. Architecture comparisons also change retention, retrieval, and context length: Section 5.1 reports approximately 217 returned words for MG versus 2,070 for LT (see Appendix J). Higher rates may therefore reflect greater access, more misuse, or both. In Appendix K.5, the agent retrieves Michelle Johnston's status and incorrectly attributes it to Michael Johnson. This illustrates misuse after retrieval, but does not establish which stage drives longitudinal or cross-architecture differences.

(2) The evidence is insufficient to establish generalizability or derive well-supported defense implications.

A. The datasets and backbone models are insufficiently diverse. Synthetic streams deliberately introduce confusable entities; the main real-data trajectory uses one Enron mailbox. Office assistants use three backbones, each of five OpenClaw models has one trajectory and a different realized workspace. Generalizability is therefore uncertain. OpenClaw's disclosure denominator also includes absent targets, potentially conflating non-retention with safe handling.

B. Practical validation and reproducibility are limited. Event-Structure has many false positives (Registrar precision: 0.404), with no intervention test demonstrating fewer violations without harming benign tasks. Reproduction also requires complete probes, prompts, seeds, configurations, and OpenClaw scripts.

C. The violation labels and trend need independent validation. The automated judge achieves precision 0.81 and recall 0.73 on an enriched 1,000-case audit (Appendix B), which does not independently confirm the increase across checkpoints.

If the paper aims to inform defenses, it must explain these phenomena rather than merely report them. Increased retrieval, cross-entity confusion, and unsupported claims may require different safeguards. Without separating their causes, the paper cannot establish where defenses should intervene.

References [1] Chen, Z., et al. AgentPoison: Red-teaming LLM Agents via Poisoning Memory or Knowledge Bases. NeurIPS, 2024.

[2] Wei, Q., et al. A-MemGuard: A Proactive Defense Framework for LLM-Based Agent Memory. arXiv:2510.02373, 2025.

[3] Mireshghallah, N., et al. CIMemories: A Compositional Benchmark for Contextual Integrity of Persistent Memory in LLMs. arXiv:2511.14937, 2025.

Questions for authors
  1. Related to Weakness 1. How is a benign interaction defined and verified? If benignity means absence of malicious intent, how can that be established in real correspondence? If defined by outcome, ostensibly benign interactions can cause the same harm as malicious ones. What operational criterion distinguishes them without relying solely on dataset origin or observed harm?

  2. Related to Weakness 1. What falsifiable failure mechanism distinguishes benign-history accumulation from established memory misuse? Could static memory compositions reproduce the findings, and what controlled experiment would rule out this explanation?

  3. Related to Weakness 1. At each checkpoint, how often do probes retrieve risk-enabling memory, and how often is that memory misused after retrieval? Can memory architectures be compared under matched retrieval budgets and retention windows, with normal task success reported alongside violation rates?

  4. Related to Weakness 2. Could human annotators label the same fixed probes at early and late checkpoints and directly verify the reported increase? How sensitive is that increase to the judge's missed violations and the strict attribution rule under which U = 0 does not necessarily mean safe behavior?

  5. Related to Weakness 2. Does the trend persist with untargeted probes, additional real histories, and repeated OpenClaw deployments? How do effect sizes change when disclosure is conditional on target presence instead of all 30 situations?

  6. Related to Weaknesses 1 and 2. How can genuine Event-Structure risk judgments be distinguished from hallucinations? At matched false-positive rates, can alert-triggered intervention reduce violations without harming benign-task success? Will the authors release prompts and configurations?

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

The paper focuses on cross-user memory contamination scenario where this memory can lead to incoreect/unsafe answers.

Critical strengths
  • Cross-user memory is that is shared among multiple users is unsafe by design. The paper focuses on this.

  • Evaluates how safety changes as memory accumulates, instead of testing a single memory snapshot.

  • Evaluates different memory mechanisms. Also on different environments and tools used models.

Critical weaknesses
  • I think cross-user memory contamination is interesting but I am struggling to understand what the realistic scenario would be in which this paper would be applicable? Such a memory system that is shared among multiple users is unsafe by design,

  • Some of the comparisons to relevant works similar to CIMemories and the paper are missing: [1] [2] .

  • The main thesis of the paper is safety failures, but sec 3.2 does include discussion around these.

  • The judge has quite a few false positives; this means that the results might be lower in reality.

[1] - "PersistBench: When Should Long-Term Memories Be Forgotten by LLMs?." ICML 2026

[2] - "GateMem: Benchmarking Memory Governance in Multi-Principal Shared-Memory Agents" arxiv 2026

Questions for authors
  • Sometimes the violation decreases. Why is this the case? There would be some confounders in this case?!

  • Is the increasing risk caused by having more memory overall, or by accumulating more potentially dangerous information?

  • Although the authors compare architectures, they do not eval whether those architectures successfully use relevant memories for legitimate tasks. This would be useful for followup works

see weakness as well

Reviewer_r97kOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper presents a trigger–probe method for testing whether fixed benign requests become less safe as an agent accumulates memory. The paired memory-free comparisons and separation of stored state, retrieval, and outcomes are strong design choices. Evidence from the OpenClaw deployment is more preliminary than the office-assistant results.

Critical strengths

Fixed probes help distinguish changing risk from simply counting more failures as interactions accumulate. Matched memory-free responses and retrieved-memory evidence make the office-assistant results more interpretable. The OpenClaw analysis tracks sensitive information spreading across files, not just whether a value remains stored.

Critical weaknesses

The human audit reports 0.81 precision and 0.73 recall for office-assistant violations. The paper reports no detectable change in error across exposure levels, but the remaining misclassification still limits confidence in the estimated rates.

Each of five model arms follows one 160-session trajectory. The observed increase in disclosures and file copies is suggestive, but a single trajectory per model cannot establish how consistently it occurs.

The office streams are synthetic or drawn from selected Enron mailboxes, and the study uses a fixed probe set. The findings do not establish risk levels for arbitrary users, tasks, or deployed memory systems.

Its high recall comes with substantial false positives, and the paper does not test whether acting on its warnings reduces violations while preserving useful task performance.

Questions for authors

Can you repeat the OpenClaw deployments with multiple independent workspace histories to test whether the disclosure and file-spread trends persist?

How do the office-assistant exposure trends change when evaluated against human-adjudicated labels or alternative judges?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.