acceptodds
Under review as a conference paper at ICLR 2027

GovernMem: Governed Evolution of Persistent Agent Memory through Failure Repair

Abstract

Long-horizon agents use persistent memory to retain experience and guide future behavior. However, memory errors can spread across interactions, and incorrect experience can also poison subsequent learning. Although prior work has advanced memory repair and experience reuse, a closed loop connecting repair outcomes and reusable experience for agent self-evolution is still missing. To address this challenge, we introduce GovernMem, a framework without updating model weights. GovernMem uses a state machine to manage repairs and collect verified feedback. GovernRouter then learns from this feedback to adjust future repair choices. As repair programs accumulate, GovernMem tracks their versions and manages their reuse. Experiments on Meta-Llama-3.1-8B-Instruct and Qwen3-14B show that matched experience reuse improves mean repair utility by 6.39 and 6.63 points. GovernRouter also achieves the highest repair utility among seven evaluated routers on both models under the matched setting.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM agents, persistent memory, memory repair, experience reuse, agent self-evolution

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_8m3xOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose GovernMem, a framework that aims to improve memory repair by learning from past experiences. A fast loop repairs memory, and a slow loop learns from this experience, and guides how memories are repaired in the future.

Critical strengths

(1) The research question is timely and relevant: Persisting errors can cause failure, and being able to reliably identify and fix memory failures could address this failure mode and meaningfully advance the state of the art.

Critical weaknesses

(1) My main concern is with the experimental setup. The authors write:

The results show that agents make better repair choices when they reuse past repair experience. Compared with starting without prior experience, using repair experience collected by another model on the same cases and in the same order improves repair utility by 6.39 points on MetaLlama-3.1-8B-Instruct and 6.63 points on Qwen3-14B.

If I understand correctly, the authors employed the same cases to gather repair experience and measure whether repair experience helps. If this is indeed the case, it would limit the applicability of the proposed solution significantly. It would be unclear if the method "recalls" specific failure cases (overfitting), or generalizes properly.

(2) The authors do not adequately discuss the limitations of their method.

Questions for authors

(1) Can the authors please clarify whether their experimental results yield any evidence how their method generalizes beyond the training manifold? It the experimental results do not yield any such evidence, could the authors please (a) include experiments that show that their method generalizes, or (b) clarify why there is a legitimate use case for their method regardless, and re-scope their writing to reflect this premise?

(2) Could the authors please share their code to ensure full reproducibility?

(3) Can the authors please include a dedicated Limitations section?

(4) Figure 3 is somewhat difficult to interpret. Could the authors enhance the visibility of the MemFail results?

3 comments

Thank you for your constructive comments. We appreciate your concerns about the experimental setup and generalization claim. We have conducted additional evaluations and will clarify the scope of our claims and improve the presentation. We believe our response has addressed most of your concerns and ask if you're comfortable to re-consider the rating in light of our efforts and dedication.


Q1: Concerns about the experimental setup and generalization.

Indeed you make it right that the original experience-reuse experiments use the same cases for source experience collection and target evaluation. This protocol measures cross-model experience reuse on shared cases. We may not have made this distinction clear enough.

So we add downstream evaluations on PersonaMem and LongMemEval-derived QA cohorts, keeping source experience frozen during target evaluation. On the 240-question controlled memory-fault evaluation, repair improves QA accuracy from 32.08% to 65.00%, a gain of 32.92 percentage points with a paired 95% confidence interval of [26.25, 39.58]. These results demonstrate downstream repair value.


Q2: Figure 3 is difficult to interpret, particularly the MemFail results.

Thank you for pointing this out. In version 1, we reduced the figure size to meet the page limit while trying to preserve the visibility of the results. We will expand the caption and accompanying discussion to explain the metrics, aggregation procedure, and MemFail results more clearly, making the figure easier to interpret.


Q3: Insufficient discussion of limitations.

We will add a dedicated Limitations section in Appendix. We will also align the abstract and conclusions with these evaluated settings, so that the scope is clear throughout the paper.


Q4: Code availability and reproducibility.

We will publicly release the code and experimental artifacts after the rebuttal period. The release will include evaluation scripts, processed evaluation data and preparation scripts, case and family manifests, configurations, random seeds, and the routing snapshots required for the reported comparisons.

We will document the commands and dependencies needed to reproduce each table, including the utility definitions, aggregation rules, and confidence-interval procedures.


We sincerely hope that these clarifications, additional downstream results, and planned revisions will help you reconsider your assessment.

Thank you for the response and the additional experiments. Unfortunately, they do not resolve my main concerns.

Generalization. The new result compares repair with no repair. The experiments do not test whether the learned experience transfers beyond the cases it was collected on. To show generalization, the experience would need to be collected on some cases and evaluated on held-out ones.

Code. The code is still not available. I can only assess what is available during the review, not planned releases.

I will therefore maintain my score.

Dear Reviewer,

Thank you again for your careful review and constructive feedback.

We have posted our responses and revised the manuscript to address the concerns you raised. As the discussion period draws to a close, we would greatly appreciate your thoughts on whether our revisions have adequately addressed your main concerns.

We would be happy to clarify any remaining points.

Thank you again for your time and consideration!

Q1: Generalization

We really apologize for not describing the transfer procedure clearly enough, which makes you confused. The downstream system was initialized with the saved GovernRouter posterior and skill library accumulated using Qwen2.5 in the original experiments. We kept this source state frozen during downstream evaluation with Qwen3. The downstream experiments therefore evaluate a system using previously accumulated repair experience.

Q2: Code availability

Our source code is now publicly available on GitHub for inspection during the review period. The current release includes the core implementation, selected evaluation scripts, and experiment manifests and does not yet include some subsequent commits from our local development branch. These files expose the repair workflow, routing updates, and experience-transfer procedure. We are organizing the remaining experimental artifacts and will add them to the repository.

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

This paper formulates memory repair into a sequential decision making problem and tries to learn good repair behaviors through experience reuse. The authors propose GovernMem, a training-free framework that combines verified memory repair with experience-driven skill selection and version management, enabling agents to improve future repair decisions by learning from past repair outcomes. Experiment across model families and data sources validate the effectiveness of GovernMem under the proposed utility.

Critical strengths
  1. The idea of formulating memory repair into a sequantial decision making problem and learning from repair experiences is interesting.
  2. The decomposition of repair utility into hierarchical residuals makes sense, and the experiments also validates its effectiveness.
  3. Experiments across model families and datasets generally show that reusing experience helps in learning good repair behaviors under the proposed proxy metrics.
Critical weaknesses
  1. The presentation is not good. Too many details are missing in the main content. See comments below.
  2. The whole evaluation relies on human-designed proxy as metrics and lacks results on realistic downstream task performances, making the evaluation not very convincing.
  3. The skill version management seems not important nor relevant to the main branch. If there is no connection between older and newer versions for a same skill, then how is it different from us treating it as a separate new skill? Do we really need the concept of version management?
Questions for authors
  1. L163: H≤tH_{\le t} not clear. What do "events" refer to here? Are they repair events?
  2. L175-176: Repeated sentence

The state machine controls memory changes and learning through five stages

  1. Can the authors formally define process faults, state drift, and poison, and show concrete examples?

Structural inconsistencies indicate process faults, verifiable replacement relations indicate state drift, and active untrusted events outside quarantine indicate poison.

  1. L194: What does this mean? Shouldn't the memory already be fixed? Does this refer to the skill version?

Commitment also requires live memory to match the starting version

  1. Eq. 2: How is bt(a)b_t(a) implemented and is it accurate or aligned with humans?
  2. L236-246: The presentation is very confusing here, with many undefined variables. Even though some details are provided in the appendix, the description in the main text is far from being sufficiently clear. Specifically, I do not understand these from main text:
  • How are Λk\Lambda_k, wi,kw_{i,k}, ρt,p\rho_{t,p}, xt,jx_{t,j}, and xtx_t implemented?
  • I guess It(a)I_t(a) should be the 1(a)1(a) in Eq. 3?
  1. The description before Sec. 3 leaves an impression that only bt(a)b_t(a) is predicted and other terms are statistical estimates, but it seems that all terms in Eq. 2 arepredicted except zt(a)z_t(a)? Correct me if I am wrong.
  2. L310: What do "source" and "target" mean here? Is it that source model generates the trajectories in the memory and target models do the repair?
  3. The definition of recurrence proxy RR does not make sense to me. It is determined by the distribution regime and repair strategy rather than observed failure recurrence.
2 comments

Thank you for your careful review and constructive suggestions. We appreciate your comments on the presentation, evaluation, and version management. We believe our response addresses most of your concerns and hope you will reconsider the rating in light of our revisions and additional efforts. Below, we respond to each point and specify the corresponding manuscript revisions.


Q1–Q2

An event is an incoming interaction or observation that triggers a memory check, but not necessarily a repair. Each repair attempt has its own transaction identifier, and its delayed outcome is processed when feedback becomes available.

We will define the external-event index tt and repair-transaction identifier ii in Section 3.1, before Eq. 1, distinguishing incoming events from outcome-settlement events.

Thank you for identifying the repeated sentence at L175–176. This editing error will be corrected.


Q3

These categories are defined through observable structural conditions:

  • Process faults: inconsistencies in projection membership, ordering, indices, permission scopes, or cache identifiers. For example, a source event exists but is missing from the retrieval index.
  • State drift: a source-derived update explicitly supersedes an earlier event, with both endpoints present and consistently ordered.
  • Poison: an untrusted audit event remains in active memory without quarantine, such as an injected instruction accessible to retrieval.

For example, if an old address is missing from the retrieval index, it is a process fault; a new address explicitly replacing the old one represents state drift; an unquarantined injected instruction constitutes poison.

We have added supplementary explanations and examples in the Appendix to clarify these faults.


Q4–Q6, Q8

Our repair mechanism follows Copy-on-Write (COW) in operating systems. It modifies an isolated shadow copy, then checks whether live memory still matches the original snapshot before commitment. If another transaction has updated live memory, the candidate is discarded to prevent overwriting newer updates.

We will clarify this in Section 3.2 ("Execution and Verification"), replacing "starting version" with "the memory-state identifier recorded before shadow execution."

For Eq. 2, the frozen backbone generates bt(a)b_t(a) for each legal candidate through structured JSON, using the diagnosis, memory summary, and candidate repair programs and contracts. Scores must be finite and within [−1,1][-1,1]. These are uncalibrated utility judgments, corrected by GovernRouter using observed repair feedback. We will clarify this process in the Appendix.

We will move essential implementation details into Section 3.3. The implementation maintains sparse Global, Pattern, and Local coordinates:

  • It(a)I_t(a) denotes active coordinates for skill revision aa.
  • ρt,p\rho_{t,p} denotes pattern responsibility (the diagnosed fault class selects one of three fixed patterns with responsibility one).
  • xtx_t denotes local features, implemented as xt=(t)x_t=(t).
  • The coordinate weight wt,kw_{t,k} is 11, ρt,p\rho_{t,p}, or ρt,pxt,j/∥xt∥1\rho_{t,p}x_{t,j}/\|x_t\|_1, depending on the level.
  • Λk\Lambda_k and ηk\eta_k are stored statistics, initialized to 11 and 00, respectively, and updated from mature receipts. They determine the correction mean μk=ηk/Λk\mu_k=\eta_k/\Lambda_k and support sk=Λk−1s_k=\Lambda_k-1.

Support thresholds are 22, 44, and 88 for Global, Pattern, and Local coordinates.

Qwen2.5-14B-Instruct serves as the source model, collecting repair experience through decisions and feedback. Meta-Llama-3.1-8B-Instruct and Qwen3-14B serve as target models, using imported experience or empty routing state. Source and target use the same repair cases in the original experiments.


Q7

Only bt(a)b_t(a) is directly generated by the backbone. The correction Δt(a)\Delta_t(a) comes from experience-based statistical estimates, while zt(a)z_t(a) is a reproducibly sampled exploration perturbation.

The final score qt(a)q_t(a) is their sum.


Q9

We agree that the implemented recurrence term is a prescribed regime- and strategy-dependent proxy, not an empirical measurement of subsequent failures.

To complement this proxy-based evaluation, we have added downstream evaluations on PersonaMem and LongMemEval-derived QA cohorts with controlled memory faults. Source experience is frozen during target evaluation.

On the 240-question controlled-fault evaluation, GovernMem improves QA accuracy from 32.08% without repair to 65.00% with repair, a 32.92-percentage-point improvement, with a paired 95% confidence interval of [26.25, 39.58].

This demonstrates the downstream task value of memory repair.


We sincerely hope these clarifications, revisions, and additional results will help you reconsider your assessment.

Reviewer_chdiOpenReview ↗

Response to Author Rebuttal

I thank the authors for their response, which partly addresses my concerns. However, I still feel the presentation and evaluation needs substantial improvement to meet the bar of acceptance, so I raise my score to 2 instead of 3.

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

This paper proposes GovernMem, a framework for continual experience accumulation and skill management through memory repair. It consists of two loops: the Fast Loop diagnoses memory failures, selects repair skills, and verifies repairs, while the Slow Loop learns from repair outcomes, costs, and recurrence to improve skill selection through GovernRouter and manage skill lifecycles. Experiments show that reusing historical repair experience improves repair quality without updating model weights or skill content.

Critical strengths
  1. Meaningful research problem. The paper addresses memory failures in long-running agents and explores how repair experience can improve future decisions, connecting memory maintenance with continual learning.
  2. Well-structured framework. The Fast Loop and Slow Loop have clear responsibilities, integrating repair verification, experience-based learning, and skill lifecycle management into a traceable improvement process.
  3. Solid evaluation of the routing mechanism. The paper compares multiple skill selection strategies and conducts ablations on hierarchical corrections, support gating, and exploration. Results show that GovernRouter effectively leverages historical experience to improve repair utility while reducing unnecessary memory modifications.
Critical weaknesses
  1. Insufficient evaluation of skill evolution. The paper proposes skill creation, revision, promotion, and retirement, but the main experiments keep skill content fixed and focus on selecting existing skills. It remains unclear whether the framework can continuously improve skills or accelerate skill evolution.
  2. Inadequate and partially unfair baseline comparisons. The paper lacks direct comparisons with skill evolution methods such as MemSkill and ERSkill. Moreover, GovernRouter benefits from pretrained routing experience, while some baselines start from scratch. The authors should better position their work within the literature and conduct fair comparisons with representative baselines.
  3. Limited experimental generalization. The benchmark contains 136 cases with three synthetically injected memory failure types. The experiments rely on three predefined patterns, while the Local component uses only event indices. Generalization to unseen failures, dynamic pattern discovery, and complex contexts remains unverified.
  4. Limited analysis of backbone capabilities. The experiments only use 8B–14B open-weight models. Evaluating stronger models, such as GPT or Claude, would help establish the framework's performance ceiling. Further analysis could identify bottlenecks in smaller models across failure diagnosis, skill generation, and selection, potentially motivating targeted training.
  5. Suggestion for an illustrative figure in the Introduction. The authors are encouraged to include a concrete example or illustrative figure to explain the two key challenges: the delayed consequences of memory repairs and the attribution of repair outcomes to evolving skill versions. This would help readers better understand the motivation behind the proposed framework.
Questions for authors

See weaknesses.

2 comments

Thank you for your detailed summary and valuable comments! We have conducted additional experiments and reconsidered the scope of our claims. We believe our response has addressed most of your concerns and ask if you're comfortable to re-consider the rating in light of our efforts and dedication. Below, we provide detailed response to all the comments.


Q1: Insufficient evaluation of skill evolution

Thank you for highlighting the distinction between experience-informed skill selection and skill-content evolution. Our experiments observed lifecycle operations such as registration, replacement, and retirement, but did not make it clear that continuously evolving skill contents improves task performance.

We will narrow our claims throughout the paper accordingly. Our claim about improvements from autonomous skill-content evolution remains outside the evaluated scope.


Q2: Inadequate and partially unfair baseline comparisons

We agree that the original comparisons did not sufficiently cover related approaches or distinguish differences in prior experience. We have added comparisons with RAG, Reflexion, ExpeL, and a QA adaptation using MemSkill’s official trained controller and skill instructions.

We have not included an ERSkill comparison because we didn't find an official evaluation protocol that we could reliably reproduce.

With a common Qwen3 backbone and retrieval evidence, the results are:

MethodQA accuracy
RAG65.00%
Reflexion63.75%
ExpeL66.25%
MemSkill skill-prior adaptation61.25%
GovernMem with the additional evidence reader67.92%

Although the multiplicity-adjusted comparisons do not establish statistical superiority, it's the highest observed accuracy among these external adaptations.


Q3: Limited experimental generalization

We have added downstream evaluations on PersonaMem and LongMemEval-derived QA cohorts with controlled memory faults. Source experience is frozen during target evaluation.

On the 240-question controlled-fault evaluation, GovernMem improves QA accuracy from 32.08% without repair to 65.00% with repair, a 32.92-percentage-point improvement. This provides evidence that repairing the memory available to the answering model has downstream task value.

Adding the evidence reader further increases system accuracy to 67.92%. This improvement is attributable to the reader extension: the experience-enabled and reset variants both achieve 67.92% with the same reader. Historical experience changes skill selection in the inspected runs.


Q4: Limited analysis of backbone capabilities

We indeed agree that evaluating stronger backbones would help characterize the framework’s capabilities. Our current experiments cover 8B–14B open-weight models under the available computational budget, so they do not establish a performance ceiling.

We will clarify this scope and discuss stronger-backbone evaluation in the next version, together with component-level analysis of diagnosis and skill selection.


Q5: Suggestion for an illustrative figure in the Introduction

Thank you for this suggestion! Actually it's really helpful it you point it out how we can make paper-presentation more beautiful. We will move our flowchart to the introduction part so that you can understand the mechanism more easily.


We sincerely hope that the additional downstream evidence and the clarified scope of our claims will help you reconsider your assessment.

Dear Reviewer,

Thank you again for your careful review and constructive feedback.

We have posted our responses and revised the manuscript to address the concerns you raised. As the discussion period draws to a close, we would greatly appreciate your thoughts on whether our revisions have adequately addressed your main concerns.

We would be happy to clarify any remaining points.

Thank you again for your time and consideration!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.