acceptodds
Under review as a conference paper at ICLR 2027

CoDiMem: Evolving Agent Memory through Intervention-Grounded Diagnostic Learning

Abstract

LLM agents increasingly rely on external memory to reuse experience across tasks, motivating recent approaches that evolve memory architectures from interaction experience. However, architecture evolution requires repeatedly deciding what to change next, while existing methods largely use previous trials to select candidate architectures rather than to improve the diagnostic policy that guides subsequent search. We introduce CoDiMem, a self-evolving agentic memory system that reuses each architecture trial in two complementary ways: as evidence for selecting the next architecture and as a factual trial record for updating diagnostic guidance, thereby allowing subsequent diagnostic decisions to build on what earlier trials actually established rather than reasoning from each new trajectory in isolation. Across AppWorld and MemoryArena with two LLM backbones, CoDiMem improves upon its paired starting architectures while reducing search cost; on AppWorld with StepFun 3.7 Flash, held-out accuracy increases from 59.03% to 69.44%. Ablations further show that removing diagnostic-guidance updates degrades both task performance and search efficiency, supporting the value of learning from accumulated architecture trials.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
LLM Agents, Agent Memory, Memory Architecture Evolution, Intervention Verification, Diagnostic Adaptation

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 3 reviews, above 27% of the others.
Reviewer_pf2yOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

CoDiMem couples memory-architecture search with updates to natural-language diagnostic guidance. An agent proposes hypotheses about memory failures, implements scoped code changes, and evaluates candidates against the incumbent through screening and confirmation stages. Trial records, including unsuccessful interventions, inform subsequent diagnosis. Experiments on AppWorld and MemoryArena with two backbones report improvements over the starting AWM architecture; on AppWorld with StepFun, accuracy increases from 59.03% to 69.44%.

Critical strengths
  1. Well-structured search procedure. Linking hypotheses, executable modifications, observed memory changes, and paired task outcomes provides a disciplined approach to architecture evolution. Retaining unsuccessful trials for subsequent diagnosis is a useful design choice.

  2. Useful empirical analysis. The paper combines held-out evaluation with search-cost accounting, a guidance-update ablation, and illustrative trial histories. These analyses provide insight beyond aggregate task accuracy.

Critical weaknesses
  1. The central ablation does not isolate diagnostic-guidance learning. Section 4.4 acknowledges that the fixed-guidance variant’s access to previous trial records and round-specific recommendations is unspecified. Consequently, the reported improvement cannot be confidently attributed to guidance revision rather than differences in available historical information. This directly weakens evidence for the central contribution.

  2. General-purpose reuse remains unestablished. The experiments demonstrate optimization within individual environments, without jointly developing one memory system across diverse benchmarks or transferring a fixed evolved architecture to an unseen environment. The results therefore leave unclear whether search discovers broadly reusable memory mechanisms or primarily environment-specific adaptations.

  3. Practical value relative to integrated memory systems is unclear. The comparisons cover several research memory methods, but do not establish an advantage over representative agents using native memory mechanisms, such as Claude Code. Given the substantial offline search cost, evidence about deployment advantages and reuse across environments would materially strengthen the practical case.

Questions for authors

Clarification: Does the fixed-guidance variant receive exactly the same historical trial records and recommendations as full CoDiMem? If so, please specify how this is ensured.

Evidence request: Have you evaluated a fixed architecture evolved on one benchmark in another environment, rebuilding its memory bank but performing no further architecture search? Alternatively, is there evidence from joint development across benchmarks?

Evidence request: Are there comparisons with integrated native-memory agents under comparable execution and resource constraints? Such evidence would help assess practical benefits beyond the research baselines.

Reviewer_t7AAOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes CoDiMem, which improves agent memory architectures by learning diagnostic guidance from previous intervention outcomes. Experiments show performance gains on AppWorld and MemoryArena.

Critical strengths

The paper addresses a meaningful problem in agent memory evolution and presents a clearly structured framework with promising empirical results.

Critical weaknesses

Weaknesses

W1. Limited experimental evaluation.
The experiments cover only two benchmarks and two backbones, with the main ablation conducted in a single setting. More comprehensive evaluations are needed to demonstrate generalizability.

W2. Insufficient ablation studies.
The fixed guidance ablation does not clearly isolate the contribution of diagnostic guidance updates from access to historical trial records. Additional controlled ablations are necessary.

W3. Limited methodological novelty.
The proposed framework closely resembles existing reflection based optimization and memory evolution methods, such as MemEvolve, ALMA, and GEPA. Its distinct methodological contribution is not sufficiently demonstrated.

W4. Insufficient analysis of diagnostic evolution.
The effectiveness of diagnostic guidance evolution is mainly supported by a single qualitative case study. Moreover, all accepted guidance updates are additions, raising questions about whether the method genuinely refines diagnostic knowledge rather than simply accumulating instructions.

W5. Insufficient manuscript content.
The main content spans only approximately eight pages, with limited space devoted to experimental analysis and discussion, leaving several important aspects insufficiently explored.

Questions for authors

My main concern has already been raised in the "weakness" section,please refer to that part.

Reviewer_QQ4LOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CoDiMem, a framework for improving agent memory architectures by learning from previous architecture-search trials. The authors argue that existing memory evolution methods use execution feedback to propose and evaluate new architectures, but do not explicitly use previous intervention outcomes to improve the diagnostic process itself. CoDiMem generates hypotheses about potential memory bottlenecks, tests targeted architectural modifications, and records the relationship between each hypothesis, intervention, and observed outcome. These trial records are then used to update persistent diagnostic guidance for future search rounds, even when the corresponding architecture changes are not adopted. Experiments on AppWorld and MemoryArena show that CoDiMem improves over its starting memory architectures. An ablation without diagnostic-guidance updates also shows lower performance and higher search cost, suggesting that learning from previous interventions can improve architecture evolution.

Critical strengths
  1. The paper studies an interesting problem in self-evolving memory systems. Rather than only improving the memory architecture, it considers how the architecture-search process itself can learn from previous attempts. This is a natural and potentially useful extension of existing memory evolution methods.

  2. I like the idea of explicitly connecting diagnostic hypotheses with architectural interventions and their outcomes. Keeping records of unsuccessful experiments is particularly useful, since a rejected modification may still provide information about which directions are worth exploring next.

  3. The architecture-search procedure is reasonably well designed. Testing targeted modifications against the same incumbent and using separate screening and confirmation stages makes the evaluation of candidate changes more systematic.

  4. The experimental results are promising. CoDiMem improves over its starting architecture across the evaluated settings, and the ablation suggests that updating diagnostic guidance can improve both final performance and search efficiency.

Critical weaknesses
  1. My main concern is that the paper does not fully isolate the contribution of diagnostic learning. The ablation without guidance updates does not clearly control for access to previous trial records. As a result, it is difficult to tell whether the improvement comes from learning better diagnostic guidance or simply from providing more historical information during architecture search.

  2. The evaluation of architecture-search stability is limited. The paper mainly reports results from individual search runs, while the final AppWorld evaluation contains only 48 distinct tasks. Since LLM-based diagnosis and architecture modification can be quite stochastic, it would be important to understand how consistent the improvements are across independent searches.

  3. The paper does not directly measure whether the diagnostic policy becomes better over time. Most of the evidence comes from the performance of the architectures discovered by the search process. While this is an important metric, it does not necessarily show that the system has learned to make more accurate or useful diagnoses.

  4. It is unclear how well the accumulated diagnostic guidance works over longer search horizons. The case study suggests that the guidance mainly grows by adding new rules. Without a mechanism to revise or remove outdated guidance, the system may eventually accumulate redundant or conflicting recommendations.

Questions for authors
  1. Could the authors compare against a baseline that has access to all previous trial records but does not update persistent diagnostic guidance? This would help determine whether the improvement comes from diagnostic learning or simply from using more historical context.

  2. Could the authors report results from multiple independent architecture-search runs? I would be interested in the variance of both final performance and search cost.

  3. Is it possible to directly evaluate the quality of diagnostic hypotheses over time? For example, does the fraction of successful interventions increase, or does the system require fewer trials to identify a useful architectural change?

  4. How does CoDiMem handle diagnostic guidance that becomes outdated or is contradicted by new evidence? Can previously learned guidance be revised or removed, and does it transfer to different starting memory architectures?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.