acceptodds
Under review as a conference paper at ICLR 2027

EvoArena: Tracking Memory Evolution for Robust LLM Agents in Dynamic Environments

Abstract

Large language model (LLM) agents perform strongly on many benchmarks, yet most evaluations assume static environments, while real-world agents must continually adapt to changing environments. To address this gap, we introduce EvoArena, a benchmark that models progressive environment evolution across terminal, software, and social domains, and EvoMem, a patch-based memory paradigm that records memory evolution as structured update histories. Current agents achieve only 39.6% average accuracy on EvoArena, while EvoMem improves EvoArena by 1.5% and standard benchmarks GAIA and LoCoMo by 6.5% and 3.3%, respectively. EvoMem further improves chain-level accuracy by 4% where success requires sustained reliability across evolving states, with analyses showing better preservation and use of evolving evidence. These results highlight the importance of explicitly modeling environmental evolution in both agent evaluation and memory. We will opensource the code and data.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
agent evaluation, agent memory, evolving environment

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

EvoArena points out that current benchmarks rely on static data, whereas in real-world environments, intermediate requirements or contextual conditions may shift, rendering existing benchmarks inadequate for accurately reflecting model performance. To address this, the authors propose a memory update mechanism (EvoMem) that preserves the differences and evidence of changes from each patch—capturing historical versions and the rationale behind modifications—and combines this with the latest memory (MTM_T) to construct a new context for the LLM.

Critical strengths
  1. This work offers a valuable insight that existing evaluation protocols have long overlooked environmental dynamics, creating a notable gap between current benchmarks and real‑world deployment scenarios.

  2. The proposed EvoArena benchmark suite covers diverse task categories across multiple domains. Meanwhile, the EvoMem memory paradigm features a clean design that is easy to understand and follow.

Critical weaknesses
  1. The environmental transitions within EvoArena are pre‑scripted and rule‑governed. Changes do not arise causally from the agent’s preceding behaviors. Although the benchmark reproduces the outcomes of real‑world environment evolution, it lacks the generative causal mechanism behind environmental shifts. This creates a non‑trivial gap compared against real‑world deployment conditions.

  2. EvoMem exhibits limited out‑of‑the‑box generalizability across different application scenarios. Its append‑only patch log generates intermediate records that grow roughly linearly with the number of agent interactions. While top‑k retrieval mitigates inference‑time context overhead, there are no native mechanisms for patch pruning, aging or compaction. Under long‑run real‑world deployment, this memory expansion poses risks of memory bloat and potential context‑related failures.

Questions for authors
  1. EvoMem records everything based on historical observations. What happens when an agent encounters a completely new environmental change that has never appeared in the past patch history? How does this patch-based memory mechanism handle such unprecedented environmental shifts?

  2. Why is the performance gain of EvoMem much larger on chain‑level accuracy than step‑level accuracy across most benchmarks? What is the core failure mode being fixed at chain level?

2 comments
AuthorsOpenReview ↗

Response to Reviewer iPuY (1)

We thank the reviewer for recognizing the importance of evaluating agents in dynamic environments, the broad coverage of EvoArena across diverse domains, and the simplicity and clarity of EvoMem's design. We are happy to address the concerns below.


W1: Environmental transitions are pre-scripted and rule-governed, which creates a gap compared to real-world deployment. Changes do not arise causally from the agent’s preceding behaviors.

We thank the reviewer for raising this point. We would like to clarify that EvoArena focuses on environment evolution driven by external factors, rather than agent-induced environmental changes. These are distinct settings, as real-world environments frequently evolve due to external factors such as software updates, policy revisions, and changing user preferences, independently of the evaluated agent's actions.

Importantly, not all EvoArena transitions are synthetic or pre-scripted. SWE-Chain-Evo is constructed from chronological commits in real-world GitHub repositories, reflecting actual software evolution. For Terminal-Bench-Evo and PersonaMem-Evo, the evolution mechanisms are grounded in realistic changes, including API and dependency updates, workflow constraints, and evolving user preferences. Successive versions are not independent snapshots. Each evolution step builds upon preceding states and accumulated histories, preserving dependencies across environmental changes.

Scripted evolution also enables controlled, reproducible evaluation across transition types, evolution stages, and chain lengths, allowing systematic assessment of agents' adaptation capabilities.

We acknowledge that agent-induced evolution represents another valuable research direction and have clarified the scope of EvoArena in the revised manuscript.


W2: EvoMem lack of patch pruning or compaction mechanisms raise concerns about memory scalability and potential context failures in long-term deployment.

We thank the reviewer for highlighting this important scalability concern. We have already investigated EvoMem's robustness to increasing history lengths in Appendix C.1. Across interaction histories ranging from 31.7K to 367.9K tokens, we find no evidence that EvoMem's relative benefit systematically declines with longer histories after controlling for task difficulty (p = 0.5673). This provides initial evidence of EvoMem's robustness under growing interaction histories.

Moreover, EvoMem does not generate patches for every interaction. As described in Section 4.2, patches are recorded only for non-additive memory revisions, while top-k retrieval bounds the number of historical patches introduced into the inference context. Appendix C.3 further quantifies the inference overhead, showing average input-token increases of only 8.8% on PersonaMem-Evo and 12.4% on LoCoMo.

EvoMem is intended to establish the effectiveness of explicitly tracking memory evolution as a lightweight augmentation to existing memory systems, rather than to provide a fully optimized memory lifecycle management framework. Our experiments demonstrate the value of this mechanism across multiple agent frameworks and memory systems, providing a foundation for further research.

We acknowledge the value of more sophisticated memory management mechanisms for long-term deployment. Nevertheless, our existing experiments demonstrate EvoMem's robustness under the evaluated range of growing interaction histories. We will emphasize this finding and encourage future work to explore mechanisms such as patch pruning, aging, and compaction. We thank the reviewer again for highlighting this valuable direction for future research.


Q1: What happens when an agent encounters a completely new environment?

This is a very interesting question. When an agent encounters a completely new environment, the initial observations typically introduce new knowledge, which is handled by the underlying memory system as additive updates. EvoMem specifically focuses on non-additive memory revisions and is designed to capture memory evolution while preserving the base memory system's additive learning capabilities, allowing agents to continue acquiring new knowledge in unfamiliar environments while benefiting from explicit tracking of subsequent memory revisions.

AuthorsOpenReview ↗

Response to Reviewer iPuY (2)

Q2: Why is performance gain on EvoMem larger on chain-level accuracy and the core failure fixed?

The larger chain-level gains arise because chain-level evaluation requires agents to adapt across successive environmental changes, during which useful historical information may be overwritten or lost as the base memory system consolidates new observations. EvoMem preserves these potentially overwritten memory states, together with their transition rationales and supporting evidence, allowing agents to recover relevant historical knowledge when adapting to subsequent changes. This is particularly beneficial for sustained chain-level performance, as reflected in Table 2.

Our diagnostic analyses in Section 6 further support this explanation across three domains:

  1. Terminal-Bench-Evo (Table 5): EvoMem's gains are larger when historical patches are retrieved (+6.5% vs. +3.1%) and incorporated into downstream reasoning (+8.3% vs. +2.6%). This suggests that recovered transition information helps agents update their reasoning and actions under changing execution requirements.

  2. SWE-Chain-Evo (Table 6): Losing previously relevant implementation information can directly lead to software regressions. EvoMem reduces the average regression rate from 9.09% to 6.32%, demonstrating better preservation of previously valid behavior.

  3. PersonaMem-Evo (Table 7): We directly evaluate historical evidence preservation and find that EvoMem improves complete evidence capture from 72.5% to 74.9%, with a +4.4% on temporal trajectory questions.

Together, these findings support our central hypothesis: preserving and recovering potentially overwritten historical information helps agents adapt more reliably across evolving environments, explaining the larger gains observed at the chain level.


Thank you again for your constructive feedback. We hope our responses have clarified EvoMem's contributions and addressed your concerns. We have revised the manuscript accordingly and would be happy to discuss any further questions.

Reviewer_Bx7YOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies LLM agents under persistent environment evolution, where workflows, software repositories, and user preferences change across related tasks. It introduces EvoArena, a benchmark comprising Terminal-Bench-Evo, SWE-Chain-Evo, and PersonaMem-Evo, and evaluates both individual-task performance and reliability across evolution chains. The paper also proposes EvoMem, a memory augmentation mechanism that preserves structured records of memory revisions, including before/after states, update rationales, and supporting evidence, and retrieves relevant revisions alongside current memory. Across the evaluated agents and models, the paper reports average improvements of 1.5 percentage points in step accuracy and 4.0 points in chain-level performance on EvoArena, with additional gains on GAIA and LoCoMo. Comparisons with alternative history representations and domain-specific analyses investigate the contributions of structured transitions, their use in agent actions, regression prevention, and preference-evidence preservation.

Critical strengths
  1. EvoArena jointly evaluates adaptation and preservation through cumulative task histories, with explicit tests for both in the software subset. This makes it valuable for assessing whether persistent agents can revise outdated assumptions without losing still-valid knowledge.
  2. EvoMem preserves past memory revisions alongside their rationales and evidence, supporting reasoning beyond the latest memory state. Its compatibility with existing agent and memory pipelines broadens the practical applicability of this approach across representations and domains.
Critical weaknesses

The paper compares EvoMem with Simple Changelog, Git-RAG, and matched-token retrieval, and additionally evaluates Mem0 with and without EvoMem. These experiments provide useful evidence for the proposed augmentation, but direct comparisons with established systems employing different memory-management mechanisms remain limited. Without such comparisons, the results demonstrate improvements over the selected host systems and history representations, but leave the relative effectiveness of patch-based memory evolution among broader memory approaches unclear.

Questions for authors

A central premise of the paper is that preserving memory transitions provides information that is lost when memory is consolidated into its latest state. Could the authors clarify whether EvoMem’s gains primarily come from recovering otherwise discarded historical evidence, or from explicitly representing the relationships between successive states? Is there evidence that the transition structure itself remains beneficial when the same historical facts and supporting evidence are available to the agent? This distinction would clarify the specific contribution of modeling memory evolution.

1 comment
AuthorsOpenReview ↗

Response to Reviewer Bx7Y

We thank the reviewer for recognizing EvoArena's value in evaluating adaptation and knowledge preservation, as well as EvoMem's evidence-grounded memory tracking and compatibility with existing agent systems. We are grateful for your positive evaluation and are happy to address the weaknesses raised below.


W1: Evaluate with broader memory systems.

Thanks for this suggestion. We specifically selected A-Mem and Mem0 as two well-established and widely recognized memory systems, representing distinct memory-management approaches: A-Mem employs an evolving memory graph, while Mem0 uses incremental memory extraction and consolidation. Evaluating both with and without EvoMem (Tables 2 and 9) demonstrates its effectiveness across different memory architectures.

More generally, EvoMem is designed as a complementary module that operates on non-additive memory revisions, without depending on a particular memory representation. This makes it adaptable, in principle, to other memory systems that expose memory updates and support additional retrieval.

Furthermore, our controlled ablations (Table 4) complement these cross-system evaluations by comparing different history-preservation mechanisms within the same host settings. This helps isolate the contribution of structured transition modeling without confounding differences between complete memory architectures.

We acknowledge that broader system-level comparisons would further establish EvoMem's relative effectiveness and have clarified it in the revised manuscript.


Q1: Clarification of the contribution.

Thank you for this question. EvoMem's benefits arise from two complementary aspects, both investigated in our experiments:

  1. Historical evidence preservation: Table 7 directly evaluates gold-evidence captured. EvoMem improves overall row-level evidence capture from 72.5% to 74.9%, with larger gains on temporal trajectory questions (+4.4%). This demonstrates that EvoMem helps preserve relevant historical evidence that might otherwise be lost during memory consolidation.

  2. Structured transition modeling: Table 4 compares EvoMem against Simple Changelog and Git-RAG, which preserve historical changes without EvoMem's structured rationales, under comparable retrieval budgets. EvoMem consistently outperforms both across all three domains, improving over Git-RAG by +3.6%, +1.0%, and +1.5%, respectively. Additionally, EvoMem outperforms Matched-token Retrieval (43.8% vs. 40.2%), suggesting that its gains extend beyond simply providing more retrieval context or history.

Together, these results demonstrate the complementary benefits of preserving historical evidence and explicitly modeling memory transitions. We thank the reviewer for this insightful question, which has helped us better clarify our contribution and its distinction from existing approaches. We have revised the manuscript accordingly.


Thank you again for your positive evaluation and constructive feedback. We hope our responses have clarified EvoMem's contributions and addressed your concerns. We have revised the manuscript accordingly and would be happy to discuss any further questions.

Reviewer_gtDwOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies a dynamic benchmark for evaluating LLM agents in evolving environments, together with a memory mechanism for adapting to such changes. Specifically, the authors argue that existing agent benchmarks mostly evaluate static environment snapshots, whereas deployed agents face environments that change across versions. To address this, they introduce EvoArena, which converts terminal, software-engineering, and personalization tasks into versioned evolution chains evaluated at both the step and chain levels. They also propose EvoMem, a plug-in memory layer that records non-additive memory updates as patches containing the before/after content, the rationale for the change, and supporting evidence, and retrieves relevant patches alongside the latest memory at inference time. Experiments across 8 backbones and 4 host agents show that current agents struggle under environment evolution, especially at the chain level, and that EvoMem consistently improves robustness on both EvoArena and standard benchmarks such as GAIA and LoCoMo.

Critical strengths
  1. The problem is interesting and under-explored.
  2. A benchmark is proposed and 8 recent backbone models across 4 agent frameworks are evaluated in this benchmark.
  3. A simple framework is proposed and the idea is straightforward and easy to understand.
Critical weaknesses
  1. The “state collapse” motivation does not hold. Specifically, the introduction section claims that “most memory-based agents maintain memory as a single latest state” and and the related work states that memory systems “usually consolidate memory toward the latest state.” However, there are many existing works already preserve superseded memory states explicitly, e.g., Zep, Git-Context-Controller, ChronoMem, DiffMem. None of these are cited or compared.
  2. The baselines do not test the claimed failure mode. The paper attributes failures to latest-state memory (“state collapse”). However, the main experiments mainly focus on the comparisons between with memory and without memory vairants. The experiments about patch-augmented memory vs. latest-state memory are required but missing.
  3. The experimental results is not solid enough to support the conclusions. For example:
  • (i) Table 2 reports raw accuracies without explaining what they indicate. There is no reference point, such as the same models on the original static tasks, and no error analysis, so it is unclear why 39.6% shows that agents “struggle under environment evolution” rather than simply facing harder tasks.
  • (ii) EvoMem’s gains are mostly a few points, several are negative (e.g., −2.4 / −2.6 on SWE-Chain-Evo, −2.9 on PersonaMem-Evo). This indicates EvoMem is not superior and is unstable.
  • (iii) Inconsistent chain metric. Chain accuracy is defined differently for SWE-Chain-Evo (fraction of consecutively solved steps) than for the other two subsets (all steps correct). However, Figure 1 averaged them together, which is not meaningful because the two definitions are on different scales and the resulting model ranking is hard to interpret.
  • (iv) There are some other issues such as some analysis only use very few model and run.
Questions for authors

Please see my weaknesses. I think this paper clearly below the acceptance threshold of ICLR. The writing should be largely polished and more experiments should be included.

3 comments
AuthorsOpenReview ↗

Response to Reviewer gtDw (1)

We thank the reviewer for the feedback and for recognizing the importance of the problem. We appreciate the concerns raised and would like to provide further clarifications, particularly regarding several points that may have been misunderstood or overlooked.


W1: "State collapse" motivation does not hold. Existing work not cited or compared

We thank the reviewer for highlighting these relevant works and will incorporate them into our discussion. We will cite and discuss all four works. We agree that our sentence is imprecise as a statement about storage: several systems keep superseded states, including Zep (invalidated edges with validity intervals), DiffMem (git history), ChronoMem (whole-memory snapshots).

The four systems differ from EvoMem in what a stored change contains and how it is used. ChronoMem reads the active version (HEAD) and searches history only to select a rollback target. GCC versions the agent's reasoning progress within a task (milestone summaries, branches), not revisions of memory about a changing environment. Zep and DiffMem do expose history at inference, but as validity-dated facts or file-level git diffs, with no record of why a revision was made. EvoMem stores each non-additive revision as one retrievable unit (before/after state, rationale, supporting evidence) and adds it to an existing host memory rather than replacing it.

Whether this difference matters is tested in Table 4. Git-RAG retrieves the same revisions as commits with metadata and diffs, i.e., DiffMem/GCC-style version history, under a comparable retrieval budget. EvoMem outperforms it in all three domains with GPT-5.5 (65.1 vs. 61.5 on Terminal-Bench-Evo, 50.9 vs. 49.9 on SWE-Chain-Evo, 43.8 vs. 42.3 on PersonaMem-Evo). We acknowledge that our original claim was overly broad and have refined it accordingly. However, this does not invalidate our motivation: preserving versioned memory alone does not ensure effective retrieval and use of historical transitions, as further supported by our experiments.

We have added the discussion of related works and clarified our novelty claims accordingly in the revision.


W2: Does not test the claimed failure mode. Mainly focuses on with and without memory. Patch-augmented memory vs. latest-state memory are required but missing

We thank the reviewer for raising this concern and would like to clarify some misunderstandings. We clarify that "latest-state" memory retrieves only current memory entries, whereas versioned memory makes historical versions retrievable at inference time.

In Tables 2–3, 20/36 configurations already augment "latest-state" memory-equipped agents (A-Mem and Memento-Skill), rather than comparing with vs. without memory. EvoMem improves A-Mem on PersonaMem-Evo (+1.7 pp step, +3.2 pp chain) and LoCoMo (+3.3 pp), and Memento-Skill on GAIA (+6.5 pp). With Mem0 (Table 9), it improves all four settings (+1.6 to +7.5 pp). Overall, EvoMem improves 22/24 "latest-state" memory-equipped configurations.

For the remaining 16/36 configurations on Terminal-Bench-Evo and SWE-Chain-Evo, whose hosts lack native memory, Table 4 additionally compares EvoMem against versioned-memory alternatives (Simple Changelog and Git-RAG). EvoMem outperforms both on Terminal-Bench-Evo (65.1 vs. 61.7/61.5) and SWE-Chain-Evo (50.9 vs. 48.1/49.9). Thus, our experiments demonstrate gains over both the latest state and versioned memory, not merely agents without memory.

AuthorsOpenReview ↗

Response to Reviewer gtDw (2)

W3.1: Accuracy not interpreted; no static reference point; no error analysis.

Each accuracy metric is already defined at the end of its construction subsection (Sections 3.1–3.3). Step accuracy measures success on an individual versioned task. Chain accuracy measures whether an agent stays correct across successive versions of the same environment. For SWE-Chain-Evo, it is the fraction of consecutive milestones solved.

Our claim is not that evolved tasks are harder than their static originals. It is that agents are unreliable when the environment keeps changing. Two results attribute these failures to evolution rather than task difficulty alone. First, some failures exist only because the environment evolves: on SWE-Chain-Evo, 9.09% of attempts break previously valid behavior (Pass-to-Pass regressions, Table 6). This error cannot occur on a single static task. Because the reference update is applied after every step (Section 3.2), these failures do not result from earlier agent errors. Second, on experiments where the host already has memory, EvoMem adds only records of how that memory was revised. It still significantly improves chain accuracy (p = 0.0285). If the failures came only from task difficulty, records of change would not help.

Also, the paper has already included failure analysis. Section 6 analyzes when EvoMem fails to help (Table 5: gains of only +3.1 pp when no patch is retrieved vs. +6.5 pp when one is), which errors it prevents (Table 6: Pass-to-Pass regressions, where a new patch breaks previously valid behavior, drop from 9.09% to 6.32%), and which evidence is lost under consolidation (Table 7, broken down by question type).


W3.2: EvoMem improvements are only a few points and occasional negative results. This indicates EvoMem is not superior and is unstable

We respectfully disagree that "a few points" and "occasional negative results" indicate EvoMem is not superior or unstable. Across the 24 model–benchmark configurations in Table 2, EvoMem improves step-level accuracy in 20/24 cases (83.3%, with one tie) and chain-level accuracy in 23/24 cases (95.8%). The average chain-level improvements are +6.1%, +2.7%, and +3.2% across the three EvoArena domains, respectively.

Moreover, EvoMem improves performance in 11/12 settings (with one tie) across standard benchmarks, achieving average gains of +6.5 pp on GAIA and +3.3 pp on LoCoMo (Table 3), demonstrating its effectiveness beyond evolving environments.

Importantly, cluster-aware paired permutation tests (Section 5.2) report p-values of 0.0125, 0.0473, and 0.0285 for the three EvoArena domains, providing statistical evidence for the observed chain-level improvements. While occasional negative results exist, the conclusion that EvoMem is not superior or unstable is not supported by the overall experimental evidence.


W3.3: Inconsistent chain-level metrics. SWE-Chain-Evo uses a different chain metric, making the aggregate rankings in Figure 1 questionable.

We acknowledge that the chain-level metrics differ across domains, but this design choice is explicitly justified in Section 3.2 and Table 2. SWE-Chain-Evo contains longer chains, where all-or-nothing accuracy would approach zero score in many settings, limiting its discriminative power. We therefore adopt the fraction of consecutively solved milestones.

However, different metric definitions do not invalidate the ranking inherently in Figure 1. Firstly, all models are evaluated on the same three benchmarks using identical domain-specific metrics and aggregation procedures. Secondly, removing SWE-Chain-Evo entirely leaves the chain-level ranking of the eight models nearly unchanged (Kendall τ = 0.96; the only change is that Qwen3.6-27B and Gemma4-31B will be tied).

Thus, the ranking provides a fair comparison under this composite evaluation. All conclusions in the paper are drawn from the per-domain results in Table 2.

AuthorsOpenReview ↗

Response to Reviewer gtDw (3)

W3.4: There are some other issues such as some analysis only use very few model and run

Our main experiments cover 8 backbone models across 3 evolving domains and 4 agent frameworks, with additional evaluations using 6 models with 2 agents on GAIA and LoCoMo (Tables 2–3), providing broad experimental coverage.

Section 6 focuses on targeted diagnostic analyses to investigate specific mechanisms, rather than replicating the full benchmark evaluation. We therefore select representative models for these analyses, while our main conclusions are supported by the broader evaluation across diverse models, agents, and benchmarks.


We hope our responses have addressed your concerns and clarified any potential misunderstandings. We have also revised the manuscript accordingly. We would be happy to discuss any further questions or concerns you may have. Thank you again for your feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.