acceptodds
Under review as a conference paper at ICLR 2027

Learning Where a Lesson Stops: Repair Boundaries for Persistent-Memory Agents

Abstract

Persistent memory allows language agents to carry lessons across episodes, but a repair learned from one failure can become harmful when reused in a nearby state that was already handled correctly. Existing methods improve what agents store and retrieve, yet the failed trajectory itself gives little evidence about where its correction ceases to be valid. We introduce RANGE, which pairs each repair with an activation support and uses matched successful trajectories as boundary evidence for behavior that should be preserved. This evidence constrains which repair–support lesson is stored and where it may act, and the learned support is enforced again at retrieval. Across ALFWorld, WebShop, and ScienceWorld, RANGE improves post-update success by 5.7–6.0 percentage points over ExpeL while reducing Regression by 45.4–50.4% relative to ExpeL.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
LLM agents, persistent memory, failure learning, scope-aware retrieval, memory management

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_q2tcOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how persistent-memory agents can reuse failure-derived lessons without disrupting behavior that was previously successful. RANGE uses matched successful trajectories to constrain each lesson’s applicability, combining memory admission, scope refinement, and scope-gated retrieval. Experiments report improved task success and reduced regression across several interactive environments.

Critical strengths
  1. The paper addresses a meaningful problem: a useful correction can become harmful when generalized beyond its appropriate context. Using successful trajectories as evidence for what should be preserved is well motivated.

  2. The evaluation covers multiple environments and considers recovery, transfer, and preservation together. The generally consistent improvements, component ablations, and sensitivity analyses provide encouraging evidence for controlled memory reuse.

Critical weaknesses
  1. The boundary judgments are insufficiently transparent. Key decisions—identifying critical actions, estimating interference, and determining unsupported scope—are described mainly through abstract controller scores. The implementation and validation of these judgments, including how applicability conditions are evaluated and narrowing is verified, remain unclear. An action appearing in a successful trajectory is not necessarily indispensable.

  2. The specific contribution of boundary learning is not fully isolated. Matched successes influence candidate generation, selection, and scope, while stronger preservation constraints also reduce memory admission. The results therefore leave uncertainty about how much improvement comes from more accurate boundaries versus better candidate generation or more conservative filtering.

  3. The abstract could be more informative.

Questions for authors
  1. Could the authors provide an end-to-end case study showing the failure, matched success, initial lesson, boundary judgment, refined conditions, and resulting behavior? A case where only the scope changes while the repair content remains fixed would be particularly informative.

  2. How frequently does support refinement rescue a candidate that would otherwise be rejected, and how much does this refinement contribute beyond admission filtering alone?

  3. What are the controller’s total model-call and token costs? Would the advantage remain against a simpler generation-and-filtering approach with comparable compute and memory admission or activation rates?

Reviewer_rkFbOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

RANGE jointly selects repairs and activation support using failures and matched successful trajectories, combining memory admission, support-only narrowing, and scope-gated retrieval. Across several agent benchmarks, it reports improved recovery and transfer with fewer regressions on nearby previously solved cases.

Critical strengths
  • The evaluation shows consistent environment-level gains across ALFWorld, WebShop, ScienceWorld, and tool-use tasks under shared actor and memory-context budgets.

  • The unscoped-retrieval ablation directly supports enforcing activation support at retrieval, while the sensitivity analysis characterizes the recovery–preservation trade-off.

Critical weaknesses
  • RANGE does not establish that its boundary witnesses identify where a repair actually lowers task success. Its verifier judges disruption of one observed successful continuation, which does not rule out another successful continuation, and Compatibility violation reuses that same verifier. Lower overall Regression therefore supports the complete method without independently validating its boundary judgments.

  • The experiments do not isolate what matched successes contribute specifically as boundary evidence. Successful examples also shape repair candidates, and increasing their number from zero to three reduces the fraction of accepted updates from 62% to 41%. The reported gains therefore do not separate more accurate activation support from better repair selection and more selective admission.

  • The evaluation does not explain how the successful archive is populated. It specifies 100 source-failure updates and read-only evaluation rollouts, with the archive initially empty, but does not describe how or when successful trajectories enter it. This leaves the experience required to supply RANGE’s boundary evidence unclear.

  • RANGE’s advantage over other selective memory-reuse mechanisms remains unestablished. MCMA selects memory abstraction according to task similarity, and MemHarness adapts retrieved guidance to the current state. Although their training regimes differ, the comparisons with ReAct, Reflexion, ExpeL, and diagnostic variants do not establish the additional value of matched-success boundary learning over these relevant alternatives.

Questions for authors

See Weakness

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

This paper introduces a new framework, Repair Activation from Neighboring Goal-Achieving Episode (RANGE), to learn the repair with an activation support during memory update. It uses the successful trajectories to estimate the boundary where the previous evidence needs to be preserved. During the retrieval, RANGE will remove lessons whose support is inactive and semantic relevance is only applied on eligible sets. The proposed RANGE method is evaluated on various long-horizon agentic tasks such as ALFWorld, WebShop, and ScienceWorld and shows consistent improvements over the baselines.

Critical strengths
  • This paper studies a key question in agents with persistent memories, during the memory update from the failure trajectories, where does the repair on the memory should stop. To address this question, RANGE pairs repair with support activation, and the same support can be used at both stages including both the learning stage and retrieval stage. The entire design is formulated in the support space to define the evidence boundary.
  • Across several benchmarks, the proposed RANGE demonstrates consistent improvements over agent memory baselines on various metrics, showing its effective in learning from the past experience.
Critical weaknesses

I wanted to apologize I wasn't able to quite understand the entire Section 3 "RANGE design" in this paper, especially on those mathematical notations.

  • My overall impression is that this paper lacks the details or clarifications connecting the RANGE design with the actual implementations. I think having a more descriptive workflow diagram or some pseudo code on the algorithm will help the audience better understand the entire design. Also it seems vague to me how the RANGE design can be integrated with the existing agentic systems to store and retrieve the memories to execute the tasks, those details are not mentioned in the paper. The authors could also consider including some examples in a case study to illustrate how the algorithm works under the hood.
Questions for authors
  • In the experiment section, the authors introduced RANGE-Stress task, but it is unclear to me how this benchmark is constructed, and why do we need a separate benchmark to evaluate the RANGE performance?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.