acceptodds
Under review as a conference paper at ICLR 2027

SCOPE BEFORE YOU PERSIST: PREVENTING CROSS-FAMILY INTERFERENCE IN AGENT MEMORY

Abstract

Persistent memory lets language-model agents improve prompts and skills without updating model weights. We show that matching retrieval scope to certification scope enables these edits to support reliable repeated adaptation across recurring task families. We study frozen-model agents on ProcStream-RSI, a 12-round code-repair stream, using Orthogonal Regression Control (ORC), an execution-grounded gate for persistent skill edits. In an intervention that holds proposals and gate decisions fixed, retrieving each accepted skill only for its originating family raises mean hidden trajectory utility from 0.713 under global memory to 0.816 and changes harmful deployments from six of eight to none. In 27 paired randomized-order streams, Scoped-ORC improves mean trajectory utility by 0.063 [0.037, 0.094] over Global-ORC, accepts 63 rather than 12 updates, and produces multiple accepted updates in 19/27 streams, with 0/63 harmful acceptances. The global control reaches 0.713, below the static agent's 0.775, because locally valid edits can interfere with unrelated families. These results establish scope matching as a complementary control for persistent agent memory: certification determines whether an edit is supported, while retrieval scope determines where that evidence authorizes its use.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
language-model agents, recursive self-improvement, continual adaptation, persistent memory, regression control, scoped retrieval, code repair

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 2 reviews, above 27% of the others.
Reviewer_YRWKOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This submission appears to be AI-generated and probably doesn't meet the quality standards for acceptance.

Critical strengths

N/A

Critical weaknesses

Likely to be AI-generated.

Questions for authors

N/A

3 comments

The reviewer states, “This submission appears to be AI-generated and probably doesn't meet the quality standards for acceptance.” We stand by the AI-use disclosure on page 11 of the submitted manuscript. If the reviewer disputes its accuracy, we ask the reviewer to specify what use was allegedly omitted and cite evidence for that claim. The quoted comment offers an impression of the prose as grounds for recommending rejection. We ask the area chair to give that inference no weight in the decision. The research community’s trust in the conference depends on reviewers exercising independent judgment and giving reasons that authors can answer and area chairs can verify.

Reviewer_YRWKOpenReview ↗

The Reviewer Has Exercised Independent Judgment and Recommends Rejection

I strongly disagree with the authors' response. My concerns are about the fundamental research quality of this submission, not merely its writing style. Even a brief examination reveals several critical weaknesses that, in my assessment, cannot be adequately addressed during the rebuttal period.

1. Inadequate benchmarking and external validation. The main experiments rely almost entirely on ProcStream-RSI, a benchmark constructed by the authors themselves. Although the paper provides implementation details and supplementary materials, its central claims remain largely unvalidated on independently established benchmarks. The only external evaluation uses a subset of just 32 HumanEval+ tasks, despite the availability of a substantially larger benchmark, and does not even evaluate the main Scoped-ORC approach. Why are there no evaluations on established agent memory or continual learning benchmarks?

2. Weak and insufficiently grounded baselines. The main comparisons rely on author-implemented variants such as Latest-only, Replay, and Self-judge, rather than strong, established methods from the literature. Although related work is cited, the experimental baselines are not adequately connected to published reference implementations or competitive existing approaches. Without such comparisons, the claimed improvements provide limited evidence of advancement over the state of the art.

3. Insufficient experimental evidence for the claimed contributions. The original ORC method fails to outperform the Static baseline, while the main improvements arise from family-scoped retrieval. The paper does not adequately establish whether its elaborate verification mechanism offers meaningful benefits beyond straightforward per-family memory management.

Beyond these experimental concerns, the manuscript requires substantial improvements in clarity, organization, and scientific presentation to meet the standards expected of an ICLR submission. I encourage the authors to examine accepted ICLR papers for examples of how to present methodological contributions and support them with rigorous, independently verifiable experiments.

To be clear, my recommendation for rejection is based on these substantive scientific weaknesses. Whether or not the manuscript was AI-assisted does not change the authors' responsibility to establish novelty, conduct convincing comparisons, and substantiate their claims.

I do not believe these fundamental shortcomings can be adequately resolved through a rebuttal. I therefore maintain my recommendation for rejection and encourage the Area Chair to evaluate the submission accordingly.

AuthorsOpenReview ↗

Reply to Reviewer: Clarification of the Expanded Evidence and Contribution

Thank you for clarifying your concerns. The revision adds experiments that directly test memory capacity, retrieval scope, and update selection.

Benchmark and external validation. ProcStream-RSI enables a controlled test of cross-family interference: the fixed-output intervention changes retrieval scope while holding proposals, acceptance decisions, and task completions fixed. The HumanEval+ experiment provides a separate transfer diagnostic. Established memory benchmarks would broaden external validation; our central evidence identifies the deployment mechanism through controlled comparisons.

Capacity-matched baselines. We added multiversion retrieval with matched storage capacity, metadata allowance, active skill length, editor, and gate. Scoped retrieval improves trajectory utility by 0.026 [0.012, 0.043] and final-bank utility by 0.060 [0.026, 0.099]. A shared-prompt-bank comparison further supports selecting the relevant skill. These experiments directly address whether additional memory capacity alone explains the gains.

Scope and selection. On the same 27 streams, Scoped-ORC exceeds Static by 0.055 [0.040, 0.071] in trajectory utility. The added gate controls separate this retrieval benefit from selection: private-only matches ORC on these streams, while the broader comparisons characterize the utility–regression trade-off. Against a scoped cumulative one-shot control with the same slot budget, sequential adaptation improves final-bank utility by 0.041 [0.023, 0.062].

These results strengthen our central contribution: identifying cross-family interference and showing how scope-matched memory preserves useful skills across recurring tasks. The assessment that scope, capacity, and selection remain experimentally unseparated does not reflect these added controls.

Thank you for considering these clarifications. We invite you to review the latest revision and its expanded experimental evidence, and to reconsider your assessment in light of the added controls and findings.

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

Self-improving LLM agents learn from LLMs feedback and update their skill permanently. However, the paper points out that this update can unintentionally harm previously learned behavior while in some other cases can improve the performance across recurring tasks. The paper proposes a continual self-improvement benchmark consisting of 9 task families across 12 adaptation rounds. Then, they introduce ORC, a skill update verification mechanism that determines if the proposed edits should be accepted into the agent’s persistent memory. The paper’s experimental results establish that persistent self-improvement can cause performance degradation and verification alone (even using their own ORC) is not enough.

Critical strengths
  1. The paper points out a specific failure mode in self-improving agents: skill edits that are beneficial for one family task can harm the performance of an agent when it is deployed on recurring task families.
  2. It introduces a benchmark for recurring task family evaluations.
  3. It also introduces ORC, a multi-channel verification reducing dependence on unreliable LLM judgments.
Critical weaknesses
  • Paper [1] given below seems to be relevant to this paper, however, it is not compared to the experimental results.
  • The baselines are only simplified or modified versions of the method rather than other related papers.
  • The paper only studies one primary frozen model and character skill, so there is no way to generalize the benefits to a larger skill library or other persistent memory mechanism.
  • The evidence for very low harmful-acceptance rates does not prove a zero failure probability.
  • The evidence showing that the ORC gate is superior compared to the simpler acceptance one is limited.

[1] Z Ma, J Liu, H Tang, Y Ma, R Wang, Y Yuan, Y Zheng, J Hao ICLR 2026 Workshop on Lifelong Agents: Learning, Aligning, Evolving, 2026

Questions for authors

Please address the weaknesses.

1 comment
AuthorsOpenReview ↗

Reply to Reviewer: Expanded Evaluation of Scope-Matched Agent Memory

Thank you for highlighting cross-family interference and the value of recurring-task evaluation. We have clarified the presentation and expanded the experiments to address your concerns.

Related work and baselines: We now discuss Ma et al.’s AgentMemoryBench and MEMs, alongside Synapse and Agent Workflow Memory. We added a workflow-inspired retrieval baseline with matched memory capacity, active skill budget, editor, and gate. Scoped retrieval improves trajectory utility by 0.026 [0.012, 0.043] and final-bank utility by 0.060 [0.026, 0.099] over this baseline.

Memory design and generalization: New controls compare scoped retrieval with a shared skill bank, learned routing, and cumulative one-shot writing. Under the same slot budget, sequential adaptation improves final-bank utility by 0.041 [0.023, 0.062] over one-shot writing. We also reran the complete editing, selection, and retrieval pipeline with GPT-OSS-120B on all 27 streams. It starts from higher baseline performance and accepts few updates; the paired trajectory-utility intervals include zero.

Harmful acceptances: We describe 0/63 as no observed harmful acceptance, not zero failure probability. We also report affected streams and a stream-level Wilson 95% interval of [0, 0.125] for observing at least one harmful acceptance.

ORC versus simpler gates: We added scoped and global Latest-only, Replay, and private-only controls. Scoped-Private reproduces all 324 ORC decisions and utility outcomes. Scoped-Replay improves trajectory utility by 0.036 [0.020, 0.053], while recording 11/287 harmful acceptances versus ORC’s 0/63. These comparisons distinguish the benefits of retrieval scope from the utility–regression trade-off of update selection.

The expanded experiments strengthen our central contribution: identifying cross-family interference and showing how scope-matched memory preserves useful skills across recurring tasks. We also revised the presentation to make the setting, method, and experimental findings easier to follow.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Response to Reviewer: Scope-Matched Memory, Expanded Controls, and Revised Claims

Thank you for recognizing the fixed-output intervention and trajectory-based evaluation. We expanded the experiments to separate retrieval scope, memory capacity, selection, and routing.

Capacity-matched memory: We added a multiversion control matching nine entries, stored-skill capacity, metadata allowance, and the 360-character active skill budget, using the same editor and gate. Scoped retrieval improves trajectory utility by 0.026 [0.012, 0.043] and final-bank utility by 0.060 [0.026, 0.099]. A shared-prompt-bank comparison also favors selecting the relevant skill. We now discuss Synapse and MEMs and specify how the lexical retrieval baseline adapts workflow-memory retrieval to our setting.

Gate contribution: We added global/scoped Latest-only, Replay, and private-only controls. Scoped-Replay exceeds Scoped-ORC by 0.036 [0.020, 0.053] in trajectory utility, but incurs 11/287 harmful acceptances versus ORC’s 0/63. Scoped-Private reproduces all 324 Scoped-ORC acceptance decisions and utility outcomes, showing that the additional channels do not alter selection on these streams. The results distinguish retrieval benefits from the utility–regression trade-off of selection.

Routing: Original runs use manifest family labels. A new online router excludes textual template/protocol cues and selects read/write slots from public I/O structure. It reaches 0.810 trajectory utility versus 0.848 with known labels; gate evaluation groups still use manifest labels. Separate fixed-memory wrong-slot testing yields final-bank utility 0.686, versus 0.764 for initial-skill fallback and 0.862 for correct retrieval. These are explicitly distinguished from the retrospective fallback mixture.

Incremental adaptation: On the same 27 streams, Scoped-ORC exceeds Static by 0.055 [0.040, 0.071] in trajectory utility. Against a capacity-matched scoped cumulative one-shot control, it improves final-checkpoint utility by 0.051 [0.026, 0.078] and final-bank utility by 0.041 [0.023, 0.062]. The 63 acceptances are first writes across family slots, supporting skill accumulation rather than repeated refinement of existing slots.

The full GPT-OSS-120B rerun starts from higher baseline performance and accepts few updates; the paired trajectory-utility intervals include zero. The expanded experiments strengthen our central contribution: identifying cross-family interference and showing how scope-matched memory preserves useful skills across recurring tasks.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.