acceptodds
Under review as a conference paper at ICLR 2027

Hurricane: Organizing Execution Experience with Evidence and Reuse Conditions

Abstract

A useful agent memory must retain the evidence and conditions that justify reusing past experience. We introduce Hurricane, which binds reusable content to source evidence, applicability conditions, dependencies, and public validation receipts. It separates retrieval relevance from reuse eligibility: an API admission gate checks retrieved records against current public state before granting access. Controlled studies examine organization, placement, and admission. Structured experience leads on HotpotQA, whereas a prose rendering of the same evidence leads on 2Wiki. Following Trace as State, prepending organized same-task state improves mean Parents exact match over appending by 16.8–25.4 percentage points across models; BFS exhibits mixed position preferences. In a constructed 36-target API cohort, Hurricane was observed to achieve the same success count as Hybrid (32/36 each) with 58.5% fewer known tokens including allocated source preparation. This single-run result is specific to that cohort. Across studies, organization, placement, and admission have distinct, task-dependent effects.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
LLM agents, agent memory, experience reuse, evidence provenance, conditional reuse, long-context reasoning

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_5nSXOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

To improve deployed agent’s performance it is becoming a common practice to augment the agent with its past experience. The key challenge is to know what experience will be beneficial for a new query/task. Prior works have focussed on relevancy of an experience, but this paper argues that relevancy is not sufficient and aims to explore mechanisms through which we can establish reusability of an experience even when it might be relevant.

Critical strengths

The paper studies an important problem of construction context for a deployed agent from its past experience. The idea to separate relevance of an experience from reusability seems useful. The authors proposal to structure the raw experience into a form that makes assessing reusability seems useful.

Critical weaknesses
  • The paper is hard to read. Terminology used does not seem to be standard and also definitions are missing before the terms are used. Especially in the abstract and the introduction ; for example prose rendering, evidence, same-task state, etc.
  • The system description in section 3 seems very abstract and hard to follow. Again there are terms that are not defined. For example line 153 – artifact bindings, host-supplied receipts, etc. A simple running example should make it easier to follow. Similarly, equation 2 and 3 in admission gate discussion, it is not clear what Rank_hybrid and Tack_B are.
  • The results are a limited set of datasets such as HotPotQA and a few others. And it appears the proposed method does not out perform the ablations studied. Additionally baselines mentioned in related works are not compared with. Agentic and reasoning tasks may be better tasks to evaluate on to see the method in action. The results are often presented without providing any insights, so it is hard to judge the significance of their results.
Questions for authors

See weaknessess above.

Also, how are the different stages of the experience curation computed, more examples on this will be useful

Reviewer_8rXBOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes Hurricane, which stores agent experience with source evidence, applicability conditions, dependencies, and validation receipts. A reuse-time gate checks retrieved records against current public state before exposing them to the agent. The evaluation studies QA organization, long-context state placement, and selective reuse on constructed API tasks. In the main API experiment, Hurricane and Hybrid both solve 32/36 targets, while Hurricane uses 58.5% fewer known model tokens including allocated source preparation.

Critical strengths

(1) Separating retrieval relevance from reuse eligibility is well motivated. Explicit conditions and validation receipts make it easier to understand why historical experience is allowed or withheld.

(2) I appreciate the content-matched QA control and the transparent reporting of mixed and negative results. These help clarify what the experiments establish.

Critical weaknesses

(1) The paper proposes a relatively comprehensive agent memory framework, but the current experiments do not sufficiently demonstrate the advantages of the complete system over existing agent memory methods.

(2) The main API experiment has only 36 constructed targets, sharing six source implementations, with one run per task and condition. This is a small and related set of tasks, so the results may depend on individual cases and model runs.

(3) The token savings may mainly reflect reduced memory input: Hurricane provides initial history on 41.7% of tasks, versus 100% for Hybrid (Table 3).

Questions for authors

Please refer to the weakness section.

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

The paper proposes Hurricane, an agent memory framework that separates the memory retrieval process and the reuse of the retrieved evidence. The framework is tested against established retrieval methods such as hybrid retrieval across three benchmarks and multiple frontier models. Hurricane achieves competitive performance yet requires considerably reduced token counts (over 60% reduction according to Table 3).

Critical strengths
  • The agent memory retrieval framework separates same- and cross-task scenarios.
  • The Hurrican framework is token-efficient (Table 3) and achieves competitive performance against more traditional retrieval methods (like hybrid retrieval with reciprocal rank fusion).
  • The benchmarks chosen for the demonstration cover both short-context (HotpotQA) and long-context (MRCR v2) settings.
Critical weaknesses
  • The evaluation employed only 36 target scenarios, which is quite a small number. There is no clear information on how these scenarios are chosen from the larger datasets used here, therefore, the results are unconvincing.
  • The performance gap between prepending and appending is large and is task- and model-dependent. There is no obvious way to determine which approach to use here for some task.
  • The paper's main innovation is moderate improvement in system design, there seems to be an apparent mismatch from the topical areas of ICLR.
  • There is no AI use disclosure paragraph required by ICLR. A large part of the paper reads like AI-written or AI-summarized content.
Questions for authors
  • What does the extra budget in "raw log + extra budget" in Table contain? After also looking at the related Appendix section, it is still unclear. Is there an example to illustrate it?
  • How often does the admission gate reject records? A rejection percentage and some explanations on why that's so are useful.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.