acceptodds
Under review as a conference paper at ICLR 2027

G-Dreamer: Long-Term Memory Evolution and Management for LLM Agents

Abstract

Large language model (LLM) agents increasingly rely on external memory to support inference-time adaptation over long-term interaction. However, accumulating more experience does not necessarily sustain memory utility. As interaction continues, useful experience can become buried in a growing memory store, while redundant, outdated, or incorrect knowledge remains available for retrieval. Consequently, memory may perform well during early accumulation yet degrade substantially in later stages, even when its overall performance remains competitive. We refer to this phenomenon as *long-term degradation*. Mitigating such degradation requires maintaining an effective memory state that continuously incorporates new evidence, remains compact enough for reliable retrieval, and revises knowledge as its validity changes. Moreover, while concrete cases and abstract patterns provide complementary forms of experience, effective compression requires preserving their provenance to avoid losing supporting evidence. We introduce **G-Dreamer**, a framework for long-term memory evolution and management for LLM agents. During wake, G-Dreamer continuously collects trajectory-grounded cases and their outcomes. In the background, sleep-time lifecycle management performs consolidation to extract reusable patterns, selective forgetting to retire redundant experience while preserving provenance, and reconsolidation to revise patterns when later evidence challenges them. Adaptive scheduling coordinates these operations according to the evolving memory state and lifecycle stage. Experiments on long-horizon interactive tasks show that G-Dreamer mitigates long-term degradation, achieving stronger performance with a smaller active memory state at little additional maintenance cost. Code: [anonymous/G-Dreamer](https://anonymous.4open.science/r/G-Dreamer-DCB6/).

Primary area:
foundation or frontier models, including LLMs
Keywords:
Agent Memory Systems, Self-Evolving agents

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

This paper introduces G-Dreamer, a long-term memory management framework for LLM agents. The method organizes memory into cases, patterns, and anchors, and introduces three lifecycle operations: consolidation, selective forgetting, and re-consolidation. The goal is to mitigate the degradation of memory utility as interactions accumulate. Experiments on ALFWorld and WebShop show that G-Dreamer can maintain competitive task performance while reducing the size of active memory.

Critical strengths

1: The paper addresses and important problem in long-term agent memory. It investigates whether stored experience remains useful as memory grows.

2: The story is well organized and the overall framework is clearly presented. I like the Introduction part with clear motivation, making the paper easy to follow.

Critical weaknesses

1: The main performance results do not fully support the paper's claims. Table 1 shows that G-Dreamer does not consistently outperform the strongest baselines. On ALFWorld, it achieves 72.7% with Qwen3-8B, below Case-pattern (74.3%), and 71.9% with Qwen3-14B, below A-MEM (75.1%). On WebShop, the improvement with Qwen3-8B is relatively small, while the Qwen3-14B result (44.7) falls substantially behind ReasoningBank (53.0). The efficiency improvements are valuable, but the results provide stronger evidence for a performance-efficiency trade-off than for consistently better task performance.

2: The evaluation setting is rather limited for a paper on long-term agent memory. First, regarding the scope and number of tasks, ALFWorld and WebShop are well-established benchmarks, but I would like to see whether the method also works on more challenging, long-horizon tasks, such as software engineering. Second, the experimental setup itself raises some concerns. On ALFWorld, the model generates a complete action sequence in a single response, which is then evaluated through symbolic replay. In addition, each task family is evaluated as a separate 500-task stream, with memory initialized from scratch. This differs from interactive agents that adjust their actions based on environmental feedback and reuse experience across different types of tasks. The WebShop experiments are also limited to only 50 tasks, which seems rather small for evaluating long-term memory.

3: The ablation study is insufficient to understand which components actually contribute to the improvements. Related to the W2, reconsolidation is a central component of the proposed framework, but the experiments do not directly test scenarios where previously valid knowledge becomes outdated or contradictory. Such experiments would help show whether the framework can actually identify and revise unreliable memories. Moreover, Table 2 only compares Case-only, Case-pattern, and the full G-Dreamer. It does not isolate the individual contributions of selective forgetting, reconsolidation, and adaptive scheduling. As a result, it is difficult to tell which components are responsible for the additional improvement over Case-pattern.

4: I'm also somewhat concerned about the robustness of the method. The framework relies on quite a few manually tuned hyperparameters, particularly for memory forgetting, pattern activation, and maintenance scheduling. It is unclear how sensitive the performance is to these choices and whether the same settings would work well on unseen tasks without further tuning. A hyperparameter sensitivity analysis would help address this concern.

Questions for authors

1: How much of the improvement comes from forgetting, reconsolidation, and adaptive scheduling individually? Would simpler memory pruning with a fixed memory budget achieve similar results?

2: How frequently are existing patterns actually revised during the experiments, and how often do these revisions improve subsequent task outcomes?

3: Can the authors evaluate memory snapshots collected at different stages on a fixed test task set?

Reviewer_W8d2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes G-Dreamer, a long-term memory management framework for LLM agents that addresses performance degradation caused by accumulating redundant or outdated memories. It combines memory consolidation, selective forgetting, reconsolidation, and adaptive scheduling within a wake-sleep framework. Experiments on ALFWorld and WebShop demonstrate improved long-term memory efficiency and competitive task performance with reduced memory size.

Critical strengths
  1. The paper highlights long-term memory degradation as an important challenge for LLM agents, showing that accumulating more experiences does not necessarily improve performance. This provides a practical motivation for maintaining memory quality rather than simply increasing memory size.

  2. G-Dreamer demonstrates a favorable performance-efficiency trade-off, reducing active memory size while maintaining competitive task performance. The long-term analysis also provides evidence that memory lifecycle management can mitigate performance degradation.

Critical weaknesses
  1. The proposed framework mainly combines existing memory management methods, including experience consolidation, selective forgetting, knowledge revision, and periodic maintenance. While their integration is reasonable, the paper can include more ablation analysis on different components.

  2. G-Dreamer does not consistently outperform existing methods. On ALFWorld, Case-pattern achieves higher success rates with Qwen3-8B, while A-MEM outperforms G-Dreamer with Qwen3-14B. These results raise concerns about whether the additional system complexity is justified by the performance gains.

  3. Although the paper compares different memory accumulation strategies, it does not adequately isolate the individual contributions of selective forgetting, reconsolidation, adaptive scheduling, and semantic capsule compression. Furthermore, experiments are limited to selected subsets of ALFWorld and WebShop, leaving the effectiveness of the framework in more diverse and realistic long-horizon environments unclear.

Questions for authors
  1. Why does G-Dreamer underperform simpler baselines on ALFWorld?

  2. Can the authors provide component-level ablations to quantify the individual contributions?

  3. Can the authors evaluate G-Dreamer on more diverse and realistic long-horizon agent benchmarks to demonstrate its generalizability?

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

G-Dreamer manages agent memory through consolidation, selective forgetting, and reconsolidation. It connects concrete cases with reusable patterns and compresses redundant cases while retaining representative evidence. Experiments on ALFWorld-derived tasks and a WebShop subset examine performance, memory size, and changes in memory utility over time.

Critical strengths

The memory design addresses a practical tension. Keeping links between patterns and representative cases allows compression without discarding all concrete support for later revision. The performance–cost trade-off is promising. On Qwen3-14B ALFWorld, G-Dreamer achieves 71.9% success with 3.73K tokens per task, compared with A-MEM’s 75.1% at 11.4K tokens. This is useful even though G-Dreamer does not achieve the highest accuracy.

Critical weaknesses
  1. The experiments do not clearly establish long-term memory degradation or show that G-Dreamer mitigates it. Figure 2 compares gains over No Memory across different task blocks in a single Heat stream. Since each block contains different tasks, a drop in gains could reflect a poorer match between stored experience and later tasks. The current analysis cannot separate this effect from a decline in memory quality. The 50-task WebShop analysis has the same limitation. Evaluating the same held-out tasks at successive memory checkpoints, with results repeated across task orders, would make this claim more convincing.

  2. The ALFWorld evaluation differs substantially from the standard interactive setting. The model generates an entire action sequence at once, and a structural judge evaluates it without step-by-step interaction with the simulator. The paper does not check whether the judge's decisions agree with execution in the official simulator. It is therefore unclear whether the reported improvements would hold under standard ALFWorld evaluation. WebShop does use the official simulator, but the experiment covers only 50 fashion-related tasks.

  3. The ablations do not show which memory maintenance mechanisms are useful. The comparison in Table 2 adds forgetting, rule revision, adaptive scheduling, repair propagation, and other changes together. It cannot tell us which of these changes matter or whether a simpler approach would work just as well. The overall gains over Case-pattern are also mixed: G-Dreamer performs worse with Qwen3-8B (72.7% vs. 74.3%) and improves the Qwen3-14B average by 1.2 percentage points (71.9% vs. 70.7%). Separate component ablations and a Case-pattern baseline with a comparable active memory budget would help establish the value of the proposed mechanisms.

Questions for authors
  1. How closely do the structural judge’s success labels agree with official ALFWorld execution on the same generated plans? Could you report the agreement rate?

  2. Does G-Dreamer’s advantage in maintaining memory utility hold across shuffled task orders and on a fixed held-out probe set evaluated at successive memory checkpoints?

  3. Which lifecycle operations account for the observed gains? Could you provide component ablations and a comparison with Case-pattern under a similar active memory budget?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.