acceptodds
Withdrawn from ICLR 2027

PagedMem: Learning to Remember at the Right Granularity for Self-Evolving Agents

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Despite the increasing reliance of LLM agents on persistent memory, whether past experience improves or degrades their performance depends on the granularity at which it is remembered. Memory that is over-broad provides no actionable guidance; memory that is over-strict constrains or even misleads the agent when the current task is inconsistent with it. Although injected text is formally soft guidance, an over-strict procedure acts as a hard behavioral prior: it suppresses exploration and triggers negative transfer when the current environment diverges from the source trajectory. Memory injection can therefore degrade rather than improve performance when the stored experience is more specific than the current execution supports, even if the memory content itself is correct. We introduce PagedMem, a self-evolving memory system that treats representation granularity and exposure granularity as separable design variables. Unlike prior methods that fix the granularity of memory at construction and inject retrieved content wholesale, PagedMem keeps multiple levels of abstraction available and progressively exposes the level that execution currently requires. It consolidates free-form reflections into semantic pages with three levels: a short header, an open-ended summary, and provenance-linked supporting records. We conduct extensive experiments on ALFWorld with three model backbones. On every backbone, PagedMem improves success and efficiency over ReAct, AWM, and MemP and substantially reduces invalid actions. These results show that representing and using memory at the right granularity can further improve agents’ ability to adapt and generalize.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Large language models, Agent memory, self-evolving Agents

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_cvUeOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

PagedMem is an external memory system for frozen LLM agents. It organizes past reflections into semantic pages containing headers, summaries, and supporting records. Memory is initially exposed at an abstract level and progressively expanded or replaced in response to execution deficits, while task feedback drives subsequent maintenance. Experiments on ALFWorld with three backbones report higher success rates and fewer invalid actions than ReAct, AWM, and MemP.

Critical strengths
  1. The paper addresses a consequential problem with a clear design rationale. Overly specific historical experience can misguide execution in a different environment. Separating memory representation from exposure granularity, and progressively revealing details, offers a plausible approach to reducing this interference. This problem formulation and design rationale are the strongest conceptual merits of the paper.

  2. The reported behavioral improvements are consistent across backbones. The complete system improves success rates and substantially reduces invalid actions on all three backbones. For example, on Qwen, invalid actions decrease from MemP’s 67/46 on Seen/Unseen to 11/4. These results suggest practical potential and make the approach worth further investigation.

Critical weaknesses
  1. The baseline comparison is insufficient. The paper compares only against ReAct, AWM, and MemP. Since ReAct has no persistent memory, this amounts to only two memory-based competitors. Directly relevant dynamic memory methods are missing, and I do not consider the current comparisons sufficient to establish PagedMem’s relative advantage. This is a major concern affecting my recommendation. Relevant additional baselines include AdaMem[1], AutoGuide[2], ExpeL[3], ReasoningBank[4] and some recent experience memory works.

  2. The ablations do not adequately validate the central mechanism. Removing hierarchical representation simultaneously changes memory content, summarization quality, and text length, preventing attribution of the performance difference specifically to granularity. Without controls such as L1-only guidance, length-matched summaries, and scheduled expansion, the independent value of execution-triggered disclosure remains unclear. This weakens the claim that granularity decoupling is the principal source of improvement.

[1]https://arxiv.org/abs/2606.05684

[2]https://arxiv.org/abs/2403.08978

[3]https://arxiv.org/abs/2308.10144

[4]https://arxiv.org/abs/2509.25140

Questions for authors
  1. Does PagedMem retain its advantage over directly relevant methods such as AdaMEM and AutoGuide when backbones, experience sources, and interaction budgets are aligned? Please provide comparative evidence.

  2. Does execution-triggered progressive disclosure provide an independent benefit over L1-only guidance, length-matched summaries, and scheduled expansion?

Reviewer_TX1SOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

PagedMem addresses the problem that past experience can mislead language agents when its level of detail does not match the current task. It organizes reflections into three-level semantic pages and progressively exposes their contents to a frozen language model in response to execution signals. Under a shared thirty-step limit on ALFWorld, it reports higher success rates and fewer invalid actions than its reproduced ReAct, AWM, and MemP baselines across three model backbones.

Critical strengths
  1. Separating memory representation from the amount of guidance exposed during execution is a useful design idea. It offers a concrete way to address overly specific historical guidance without modifying the underlying language model.

  2. The reported success-rate and invalid-action improvements are consistent across all three evaluated backbones. The ablations and trajectory examples also help explain why hierarchical guidance and progressive exposure are useful in the tested setting.

Critical weaknesses
  1. The mechanism attribution is incomplete. The flat-memory ablation changes both organization and the injected material, while the one-shot ablation does not separate execution-conditioned timing from prompt-length and detail effects (Section 4.3). Algorithm 1 supports replacement and an L1 reset, but the operational criterion for distinguishing page mismatch from insufficient detail needs clarification. Transition-level outcomes would help establish when escalation or replacement is beneficial.

  2. The efficiency and robustness evidence is limited. Environment steps exclude reflection, organization, and other memory-management computation, and unconditional step averages mix success with failures that exhaust the budget. In Table 1, the DeepSeek Unseen improvement over MemP is 121/134 versus 119/134, or two tasks; paired uncertainty and sensitivity to online task order are not reported. This does not negate the larger gains elsewhere, but limits the interpretation of this comparison.

Questions for authors
  1. Additional validation: Can the organization and exposure comparisons use the same reflection records, a fixed detector, and matched or reported token budgets? Please also specify the mismatch criterion and report outcomes after escalation versus page replacement.

  2. Reporting and robustness: Can the authors report policy and memory-management calls/tokens, matched-task step comparisons, and paired per-task outcomes? Repeated runs or task-stream permutations would help assess the small DeepSeek Unseen gain under online memory updates.

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

The paper introduces a self-evolving memory system for long-horizon agent PAGEDMEM. PAGEDMEM gets rid of fixed granularity of memory at construction and inject retrieved content wholesale, PAGEDMEM keeps multiple levels of abstraction available and progressively exposes the level that execution currently requires. It consolidates free-form reflections into semantic pages with three levels: a short header, an open-ended summary, and provenance-linked supporting records. On a high level it has 3 layers: header, summary and supporting record. It's validated in ALFWorld Seen and Unseen with 30-step interaction budget tested on 3 backbones (Qwen3.5-35B-A3B, DeepSeek-V4-Flash, GPT-5.6-Luna) vs other memory frameworks including ReAct, AWM and MemP. PAGEDMEM shows better performance .On Qwen3.5-35B-A3B, PagedMem obtains the highest success on both splits: it improves over ReAct by 14.29 points on Seen and 14.93 points on Unseen, and over MemP by 6.43 and 5.97 points. Its combined success is 242/274 (88.32%), versus 225/274 (82.12%) for MemP and 202/274 (73.72%) for both ReAct and AWM. Across the three backbones, PAGEDMEM attains the highest success on both splits. AWM improves over ReAct on Seen (+2.14) but underperforms on Unseen (−2.24), indicating that its gains are not uniform across splits. PAGEDMEM also uses the fewest mean and median steps on both splits, reducing mean steps relative to MemP by 0.52 and 1.32.

Critical strengths

1, well-defined problems with clear solution. The paper separates how memory is stored from how much is exposed, and defined 3 layers of structure for self-evolving agent system and tested across multiple backbones. 2, the performance is consistent across all the backbone LLMs

Critical weaknesses

1, the design is interesting but carries heuristic assumptions that are not data-driven, which I think is less informative 2, the benchmark is limited and may carry some bias, more benchmarks can be tested 3, more ablation can be done for execution-conditioned, coarse-to-fine page in can be done to understand where the win comes from 4, how the thresholds were picked for header, summary are not clear

Questions for authors

1, please add baselines from Reflexion, ExpeL or ReasoningBank style under the same protocol. 2, add more benchmarks with a different action-validity structure, e.g. WebShop, ScienceWorld or AppWorld.

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

Related papers

Discussion (0)

Sign in to comment.