acceptodds
Desk rejected at ICLR 2027

Why Do Self-Evolving LLM Agents Fail? A Study of Memory-Based Learning at Test Time

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Memory-based self-evolving LLM agents promise continuous improvement by reusing experience without updating model weights. Yet a failure can originate in earlier learning, memory retrieval, or task execution, and outcomes alone cannot tell which. We introduce WTF (What's the Failure), a human-calibrated framework that diagnoses failures by linking task trajectories to memory artifacts and update histories, and tests selected diagnoses by replaying tasks from saved memory state. Across 9,544 traces spanning four self-evolution methods, five benchmarks, and three base LLMs, we derive a taxonomy of 25 failure classes. Most failures reflect persistent model, environment, or harness limitations, but 10.9% are memory/state-induced: a *memory tax*. This tax is heaviest in methods that inject their entire memory into every task, which show nearly three times the memory-induced share of methods that retrieve a few entries, and the gap widens as memory grows. Lessons from successful attempts can discourage useful actions, and correct knowledge can be missed at retrieval or applied beyond its scope. In controlled reruns, removing harmful guidance restores useful behavior without adding knowledge; across five selected cases, a memory-free agent raises pooled success from 1/18 to 13/18. Yet fixing a targeted behavior does not guarantee task success, and a fix for one task may not transfer. Reliable self-evolution thus requires effective learning, appropriate retrieval, grounded execution, and trustworthy evaluation. We will release code, memory artifacts, and task traces.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM Agents, Recursive Self-Improvement, Self-Evolving Agents, Agent Memory, Failure Analysis, Test-Time Learning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_ssBqOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces WTF, a framework for diagnosing failures in memory-based self-evolving LLM agents using execution traces and memory histories. It analyzes 9,544 traces across four methods, five benchmarks, and three models, producing a taxonomy of 25 failure classes. Selected replay experiments demonstrate that removing harmful memory can improve task success, while the broader analysis associates whole-store memory injection with more memory-related failures.

Critical strengths
  • Linking task traces to historical memory snapshots is valuable. The example in Appendix E.7 demonstrates how ignoring update timing can produce an incorrect diagnosis.

  • The controlled replays provide useful evidence for specific harmful-memory mechanisms. Separating behavioral recovery from full task success in Table 2 also makes the intervention results more informative.

Critical weaknesses
  • Diagnostic accuracy is insufficiently validated. Human calibration is useful, but no quantitative pre-reconciliation agreement or held-out diagnosis accuracy is reported. Appendix C only reclassifies existing diagnoses, leaving errors in the underlying explanations untested.

  • The delivery-mode comparison is confounded. Section 6 compares different agent systems rather than isolating memory delivery. Update failures further complicate the comparison: eight AWM runs on DeepSeek never stored a workflow, potentially reducing their exposure to harmful memory.

  • The memory-growth conclusion is not well isolated. Appendix F.1 uses fixed task orders, confounding stream position with task identity. Comparing trends across methods helps, but does not rule out method-specific difficulties on later tasks.

  • The replay evidence has limited coverage. The 14 memory-free comparisons support selected mechanisms, but cases were selected partly for reproducible failures. They cannot establish how reliably the diagnostic framework identifies memory-induced failures across the corpus.

Questions for authors
  • From the existing human annotations, can you report agreement before reconciliation, particularly on memory-induced versus base-model failures? Were any diagnoses independently evaluated on traces excluded from calibration and taxonomy construction?

  • Using the existing logs, do the delivery-mode gap and stream-position trends persist on matched task/model subsets after excluding runs with failed memory updates or empty memory? This would help assess how much of the reported difference is attributable to memory delivery.

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

Memory-based self-evolving LLM agents promise continuous improvement by reusing experience without updating model weights. Yet a failure can originate in earlier learning, memory retrieval, or task execution, and outcomes alone cannot tell which. This paper introduces WTF (What’s the Failure), a human-calibrated framework that diagnoses failures by linking task trajectories to memory artifacts and update histories, and tests selected diagnoses by replaying tasks from saved memory state.

This study uses close to 10K traces spanning four self-evolution methods, five benchmarks, and three base LLMs to derive a taxonomy of 25 failure classes. It shows that most failures reflect persistent model, environment, or harness limitations, but 10.9% are memory/state-induced, hence called memory tax. This tax is heaviest in methods that inject their entire memory into every task. Lessons from successful attempts can discourage useful actions, and correct knowledge can be missed at retrieval or applied beyond its scope. In controlled reruns, removing harmful guidance restores useful behavior without adding knowledge; across five selected cases, a memory-free agent raises pooled success from 1/18 to 13/18. Yet fixing a targeted behavior does not guarantee task success, and a fix for one task may not transfer. It concludes that reliable self-evolution requires effective learning, appropriate retrieval, grounded execution, and trustworthy evaluation.

I feel the overall performance and experiment study is interesting. It points our some current implementations of memory mechanism may not benefit the retrieval and LLM-based agent tasks. However, the reviewer believe instead of blaming “memory tax”, one should consider how to improve or redo the current design and implementation of some agent memory mechanism.

Critical strengths

S1. This is an interesting and insightful empirical study of how persistent memory can contribute to LLM-agent failures. It carefully traces failures across memory creation, updates, retrieval, and execution. It demonstrates some concrete examples, conducts some careful analysis step-by-step to illustrate its limitations. The empirical study looks very good.

S2. The case studies are concrete and in some detail and thus convincing. The examples on payment-cards, telecom turn-budget, and gift-wrap recipe show pathways from experience reuse which leads to wrong or inefficient actions.

S3. It points out correctly that the memory tax is heaviest in methods that inject their entire memory into every task, which show nearly three times the memory-induced share of methods that retrieve a few entries, and the gap widens as memory grows.

Critical weaknesses

W1. Although using some good experiments, the study points out some current implementations of agent memory mechanisms may lead to decrease the performance and hence coined a term, “memory tax”, it is hard to be generalized to other agent memory design/implementations. Thus, the blame should only be confined to some not so-well developed, current agent memory implementation methods instead of blame agent memory itself. The memory tax may only be applicable to some poor design but cannot be generalized to other design/implementations. Philosophically, agent memory should benefit more than not using memory. I believe one should have a thorough discussion once seeing such a phenomenon to point out how to improve instead of quickly generalizing it to all the cases.

W2. Following this philosophy, I feel that claim on memory tax may need to be more precise on what conditions it may lead to weaker performance in comparison to no-memory at all. The condition and boundary should be better defined so the readers may not get misconception that it is a general case and there is no way to fix it.

W3. It seems the paper is mainly an empirical study and it shows some negative results of current agent memory design/implementation. However, a more balanced discussion to work out the boundaries and cases may give readers a more balanced picture rather than a general negative view on the role of agent memory.

W4. It is good to point out some design problems as shown in this paper but it could be more useful to give people insight on how to overcome such problems. Such a discussion could enhance the paper and point out directions for future study.

Questions for authors

I would like to see how the authors would address my concerns raised in the weakness. I am open for in-depth discussion of the paper.

1 comment

There is no rebuttal submitted so far from the authors. I will maintain my score as "weak rejection".

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

This paper proposes WTF, an LLM based framework for diagnosing failures in self-evolving agents. By analyzing 9,544 task traces, the authors identify 25 failure categories and investigate how accumulated memory can negatively affect agent performance through selected counterfactual interventions.

Critical strengths

The paper addresses a relevant research topic and is generally well structured and easy to follow. The authors provide detailed descriptions of the experimental setup and illustrative case studies, which help readers understand the observed failure behaviors in self-evolving agents.

Critical weaknesses

Weaknesses

W1. Limited Novel Scientific Insights

Most identified failure phenomena overlap with prior work on harmful memories, negative transfer, and retrieval failures. Although the paper provides a broader empirical characterization, it remains unclear what fundamentally new or generalizable scientific insights emerge from this analysis.

W2. Limited Methodological Novelty and Missing Baselines

WTF largely combines existing techniques for trace inspection, LLM-based diagnosis, taxonomy construction, and counterfactual replay. The paper lacks comparisons or ablations demonstrating that its key components provide measurable advantages over simpler failure analysis baselines.

W3. Insufficient Validation of Causal Diagnoses

The main quantitative findings depend on automated causal diagnoses whose accuracy is not independently evaluated on a held-out human annotated test set. Cross model agreement in categorizing generated diagnoses does not establish the correctness of their underlying causal explanations.

W4. Limited Causal Evidence for General Claims

While the selected counterfactual interventions support individual failure mechanisms, they do not establish the population-level prevalence or average impact of the memory tax. Comparisons between full memory injection and selective retrieval remain observational and potentially confounded.

W5. Selection Bias and Limited External Validity

The counterfactual experiments focus on selected, reproducible failures, limiting the generalizability of the reported improvements. Moreover, the proposed taxonomy lacks sufficient independent evaluation on unseen agent architectures or task distributions to establish its broader diagnostic utility.

Questions for authors

My issue has already been presented under "weakness",please refer to the "weakness" section.

Comments on the paper

1 comment

Dear Reviewers and Area Chair,

Thank you for your time and your thoughtful feedback! We have decided not to submit a rebuttal at this stage. We will use your comments and suggestions to improve the next version of this work.

Sincerely, Authors

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

Related papers

Discussion (0)

Sign in to comment.