acceptodds
Under review as a conference paper at ICLR 2027

Retrieve and Compress: Why Naive Experience Sharing Fails for Web Agents

Abstract

Giving a web agent access to other agents' past trajectories is widely assumed to help. We show that unselective sharing does the opposite. On WebArena-Verified (812 tasks, 5 seeds, one independent memory pool per seed), injecting every accumulated trajectory into every subsequent task drops success from 30.0% to 12.6%, below an agent with no memory at all, as does a curated pool read at random (−9.9 pp) — in neither case is what gets injected chosen for the task in front of it. What recovers the gain is not sharing but discipline: deciding what may enter the pool, what survives compression, and which entry a task is shown. That reaches 41.4% against a 30.0% baseline (+11.4 pp, task-level paired p < 0.005, Holm-adjusted across the 13 memory conditions) — real but modest, Cohen's h = 0.24. Of those three decisions only one changes the sign: that design's own pool, read without the intent rule, scores 20.1% — 21.3 pp lower, below the no-memory baseline — while removing compression or admission leaves it above. The direction holds across task type, task difficulty and model family, and more weakly on a second benchmark (transfer results in the appendix). An oracle injecting each task's best historical trajectory reaches 53.4%, 12.0 pp above us — an upper bound, not headroom — and no configuration of ours approaches the strongest specialised systems on the benchmark.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM Agents, Web Agents, Experience Reuse, Memory, Retrieval, WebArena, Empirical Study

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_67dsOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies why sharing past trajectories can make web agents perform worse than using no memory. The authors combine filtering successful trajectories, compressing the injected text, and retrieving experiences based on task intent, and use comparisons and ablations to show that selecting experiences matters more than simply accumulating them.

Critical strengths

The clearest strength is the extensive experimental coverage, including the full task set, multiple seeds, random and failed-memory controls, and checks with a second model and an offline environment.

Critical weaknesses

Filtering experiences, shortening memory text, and retrieving relevant trajectories are common choices in existing agent memory methods, and their combination here does not establish a clearly different approach. More importantly, the finding that irrelevant experiences can hurt while relevant ones help does not substantially advance the understanding provided by prior work. Meanwhile, this paper appears to bear strong traces of AI‑generated content, with non‑standard phrasing throughout the main text.

Questions for authors
  1. Which validated finding changes our understanding of experience reuse beyond AWM, ReasoningBank, and prior memory-management studies?
  2. Does the learned VAE index still outperform raw embeddings or random projection when the memory pool, deduplication rule, and injected text are held fixed? It is necessary to present controlled results demonstrating an independent and consistent benefit from the learned index.
  3. Under the definition in A.2, why is the relevant-entry rate 78% while Recall@3, defined as retrieving at least one relevant experience, is only 64%? It would be better to have an explanation regarding this point.
  4. Does the full method retain a meaningful benefit when historical experiences from tasks with the same official template are excluded?
Reviewer_txqaOpenReview ↗
Rating
2
Confidence
3
Informativeness
3(+1)
Presentation
2
Soundness
2
Read review
Paper summary

This work studies shared experience memory for web agents: whether agents benefit from reading a pool of other agents' past trajectories.

The central hypothesis is that such memory helps only when it is curated, compressed and selectively retrieved. Under this hypothesis, injecting everything is harmful, and selection, rather than admission or compression, determines whether the effect is positive or negative.

The claim is supported along several experimental directions:

  • a multi-condition comparison on WebArena-Verified with component ablations and length controls;
  • re-implemented memory baselines;
  • a limited cross-model check;
  • an offline transfer check.

Notable aspects of the experiments: negative controls show that unselected or wrong memory hurts, while length-matched irrelevant text does not, and the statistics are paired and carefully reported.

Critical strengths
  1. Research significance and relevance: The paper addresses a timely and important problem in LLM-based agents: how to effectively share and reuse past experiences across agents. As agents increasingly operate in long-running and collaborative settings, understanding when shared memory helps or hurts performance is essential for building reliable and scalable agent systems. The paper's investigation of template reuse and retrieval quality provides useful insights into this problem.

  2. Careful paired statistics and informative negative controls: task-level McNemar tests with the full Holm family printed, a length-matched Filler control, and controls that inject random or failed memories through the same pipeline (§4.6, §5.4–5.5, Table 2).

Critical weaknesses
  1. The evidence is broad within one benchmark/pipeline, but narrow across benchmarks and model families.
  • The evidence supports compact representation/payload design more strongly than it supports the particular learned compression mechanism. An untrained random projection already reaches 35.9% (Table 1), and every width/β/threshold setting lands within ±3 pp (Fig. 5).
  • The main result is modest: +11.4 pp over a deliberately unoptimized scaffold, which the authors themselves call a small effect (Cohen's h = 0.24, §5.3). The full design (41.4%) is below the no-memory AgentOccam scaffold (43.1%; Table 12, on a different evaluator). The negative finding that injecting everything hurts is close to expected, and §2.2 already lists recent systems that retrieve selectively. All baselines are the authors' own re-implementations, and the Synapse port is described as a floor (§4.4).
  • The experimental evidence is thin. There is one interactive benchmark (WebArena-Verified), and one model (DeepSeek V4-Flash) runs every condition. The second model (GPT-5-mini) runs only A and D on the 258-task hard subset (Table 4). The "second environment" is an offline Mind2Web replay with 3 seeds and only A/B/D (Table 13). The third-model run is a 12-task pilot (36 runs per condition; C.1). The pool-size, retrieval-size, and hyperparameter sweeps are single runs with no replicates (Tables 10–11, Fig. 5). The 16 conditions are configuration variants of a single pipeline.
  • The paper needs further fine-grained revision: In the main text, the only results are Table 1 and Figs. 1–3. Tables 2–13 and Figs. 4–5 are all in the appendix.
  1. The layout and organization are extremely poor (details in the Presentation field).
  • The related work section should be further compressed and reorganized.
  • The main-text evidence is crammed onto p.6 while §5 relies on appendix tables. At least Fig 4 & 5 can be reformed into the main text after reasonable context reform.
  • Several cross-references are broken ("Appendix 8", "Appendix 4", "Section 4.6", "condition G of Table 7").
  • C.7 is an empty heading. Figures and tables float into the wrong appendix sections (Fig. 5 into C.9; Tables 4–5 into the middle of Appendix B's numbered list).
  • Fig. 2 leaves most of its points unlabeled.
  • Task categories and conditions change names between text, tables, and figures ("mutation"/"computation"; C as "curated pool"/"random store"/"random select"; "Noadm"/"Randpr").
  • The headline claims carry sign errors ("costs +21.3 pp").

Together, these make the paper hard to read and its results hard to check.

Questions for authors

Main questions:

Q1. What is D's gain over A when retrieval also excludes entries from the querying task's intent template? Without this, the +11.4 pp cannot be separated from template reuse.

Q2. What is an "episode" on the x-axes of Fig. 1 and Fig. 4, which stop at 60, when each seed's stream covers all 812 tasks and three agents? How large is the pool when Table 5's "steady state" is measured?

Minor questions:

  • Is the 374-task slice "mutation" or "computation"? What are the benchmark's actual category definitions?
  • How often is the per-seed VAE retrained as the pool grows, and what function maps latent width to the payload character budget?
  • Why does GPT-5-mini run only conditions A and D, and only on the hard subset?
  • Which conditions are the unlabeled points in Fig. 2?

Suggestions (future work; these did not affect my score):

  • Run a rather comparable, interactive second benchmark and at least one more backbone on the full task set.
  • Move the key ablation and cost tables into the main text, fix the cross-references, and use one name per condition throughout.
Reviewer_buZfOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies how having access to past trajectories can assist web agents. The paper finds that naively adding the past experiences into the context hurts and summarization and selection must be done for experiences to be helpful.

Critical strengths
  • The paper reports all experiments and results in detail and uses careful controls.
Critical weaknesses
  • The paper is hard to read and shows sign of heavy AI-generated prose.
  • The findings echo existing work[1]. Most recent work use selection and curation by default.
  • Methods like AWM, ReasoningBank, Contextual Experience Replay are cited but not compared against. The compared baselines are older and weaker.

[1] Zidi Xiong, Yuping Lin, Wenya Xie, Pengfei He, Zirui Liu, Jiliang Tang, Himabindu Lakkaraju, and Zhen Xiang. How Memory Management Impacts LLM Agents: An Empirical Study of Experience-Following Behavior. ACL 2026.

Questions for authors
  • How do the findings differ from Xiong et al. (ACL 2026)?
  • Why are AWM and ReasoningBank not run as baselines? The paper discusses both as the closest work, and both report WebArena results.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.