acceptodds
Under review as a conference paper at ICLR 2027

Memory Bank of Finite Capacity for LLMs based Multi-Agent Systems

Abstract

Memory mechanisms are the core enabler for large language model based multi-agent systems (MAS) in downstream tasks. Existing works enhance the effectiveness of memory guided MAS through the integration and updating of memory for adaptation to constantly evolving tasks. Nevertheless, these strategies incur higher consumption of storage space, time, and token resources, which constrains the efficacy of memory mechanisms in long-term MAS operation. To solve these drawbacks, this paper proposes FiniteMem, a finite memory mechanism based on multi-dimensional memory replay for MAS. It builds a three-tier memory architecture to supply both macro and micro guidance, and designs the memory replay based on memory competition and inhibitory control, to substantially reduce resource consumption while improve the memory mechanism adaptability. Experiments on multiple publicly available datasets and MAS frameworks verify that FiniteMem achieves state-of-the-art performance, reducing storage space by 31.3%, token consumption by 26.9%, and runtime by 11.8%. FiniteMem offers a universal, high efficiency long-term memory management solution for balancing performance and resource overhead of MAS. Our code will be available at https://anonymous.4open.science/r/FiniteMem-5A3B.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Multi-Agent Systems, Memory Mechanisms, Large Language Models, Long-Term Memory

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_Z362OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes FiniteMem, a finite-capacity memory management framework for LLM-based multi-agent systems (MAS), aiming to improve task performance while reducing storage, token consumption, and runtime overhead. The framework introduces a three-layer memory architecture consisting of trajectory, knowledge, and perspicacity layers, providing both fine-grained action guidance and high-level reasoning support. It further develops two memory replay strategies based on memory competition and inhibitory control to selectively retain valuable information under capacity constraints. The authors evaluate FiniteMem across five benchmarks and three multi-agent frameworks (AutoGen, DyLAN, and MacNet), reporting consistent performance improvements over existing memory methods, alongside reductions of 31.3% in storage, 26.9% in token consumption, and 11.8% in runtime compared with G-Memory. Additional ablation studies, sensitivity analyses, and experiments with smaller language models further support the effectiveness of the proposed approach.

Critical strengths
  • FiniteMem addresses the important challenge of managing growing memory in long-running multi-agent systems. It achieves competitive or better task performance while reducing storage, token consumption, and runtime, demonstrating practical value beyond accuracy improvements alone.

  • The method is evaluated across five benchmarks and three multi-agent frameworks, with additional experiments using different model sizes. The ablation studies and sensitivity analyses further support the effectiveness of the proposed memory architecture and replay strategies, suggesting that the approach can generalize across different settings.

Critical weaknesses
  • The paper compares FiniteMem mainly against methods without hard memory capacity constraints, but lacks comparisons with simpler eviction strategies under the same memory budget. Moreover, the main experimental capacity limits scale with dataset size, leaving the claimed constant-space advantage insufficiently established. These issues make it difficult to isolate the benefits of the proposed replay algorithms.

  • Appendix B.3 describes an AutoGen configuration in which a Ground Truth Agent checks outputs against standard answers. Are these answers accessible during evaluation or only used for feedback on completed tasks? This distinction is important because access to ground-truth answers during evaluation could compromise the validity of the reported results.

Questions for authors
  • Appendix B.3 states that the Ground Truth Agent checks outputs against standard answers, while Section 3.1 indicates that only correctly answered trajectories are stored. Could the authors clarify when and how these correctness signals are obtained? In particular, are ground-truth answers accessible during test-time execution, or only after a task has been completed? This distinction is important for assessing the validity of the evaluation.

  • Could the authors compare FiniteMem with simpler memory retention strategies, such as FIFO, random eviction, or recency-based selection, under the same memory capacity? Such comparisons would help establish whether the improvements come from the proposed replay mechanisms rather than memory capacity constraints alone.

  • The main experiments set memory capacities proportional to dataset size, while the paper emphasizes constant storage complexity. Could the authors clarify how these capacity limits would behave in an open-ended task stream where the total number of future tasks is unknown?

Reviewer_K75fOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces FiniteMem, a bounded-capacity memory mechanism for LLM-based multi-agent systems. FiniteMem organizes past successful multi-agent trajectories into three layers: (i) a trajectory graph that stores agent-level steps and cross-trajectory relations, (ii) a perspicacity layer containing abstract reasoning strategies and lessons, and (iii) a knowledge layer containing extracted domain knowledge. To keep these banks within prescribed capacities, the trajectory layer uses a greedy replay procedure based on intra-cluster diversity, inter-cluster diversity, and graph-derived importance, while the two abstract layers use an inhibitory-control score combining temporal decay, retrieval frequency, and a validity score. At inference time, the system retrieves high-level perspicacity and knowledge as well as successors of related trajectory nodes to guide the agents.

The method is evaluated with Qwen3-32B on five QA/fact-verification benchmarks and three multi-agent frameworks, against seven adapted memory baselines. The paper reports modest average accuracy improvements over the strongest multi-agent baselines and reductions in memory storage, token use, and runtime. Ablations remove individual memory layers or their replay mechanisms, and the appendix includes capacity, retrieval, and hyperparameter sensitivity studies as well as results using smaller backbone models.

Critical strengths
  1. The paper targets a consequential systems problem. External memories that grow with every interaction eventually become expensive to store, search, and place in prompts. A bounded memory mechanism that preserves useful experience across long-running multi-agent systems would have clear practical value.

  2. The proposed architecture considers memory at multiple semantic levels and is tested across several settings. Separating action-level trajectories from abstract reasoning strategies and domain knowledge is a reasonable design, and evaluating the same mechanism across five datasets, three multi-agent frameworks, and multiple model sizes is broader than a single-task demonstration.

Critical weaknesses
  1. The evaluation protocol is unclear and may expose test labels. The method stores only successful trajectories and uses a “Ground Truth Agent,” but the paper does not explain when gold answers are revealed, when predictions are scored, or when memory is updated. A strictly held-out, causal evaluation is needed.

  2. The capacity and efficiency claims are not supported by consistent experiments. The paper alternates between dataset-dependent and fixed capacities, reports mainly relative costs, and omits uncertainty estimates. It should report results under genuinely fixed capacities with absolute storage, token, and runtime measurements against matched baselines.

  3. The baseline comparison and statistical reporting are incomplete. Several baselines are adapted from single-agent settings, but the paper provides limited evidence that these adaptations are equally optimized. The relatively small accuracy gains should be accompanied by per-run results, variance, and significance tests.

Questions for authors
  1. When are gold answers revealed, predictions scored, and memories updated? Please report results on a strictly held-out test set with no label-derived memory.

  2. Which capacity setting produces the main accuracy and efficiency results? Please provide fixed-capacity results with absolute storage, token, and runtime costs.

  3. How were the single-agent baselines adapted and tuned for each MAS? Please report variance or significance tests for the main comparisons.

Reviewer_wrR7OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes FiniteMem, a three-layer memory system for multi-agent systems, with capacity-constrained selection policies for trajectory, reasoning, and knowledge memories.

Critical strengths
  1. Addresses the practical problem of memory growth and resource consumption.
  2. Evaluates multiple MAS frameworks and includes component ablations, capacity analysis, and runtime measurements.
Critical weaknesses
  1. G-Memory already combines hierarchical memory with high-level insights and fine-grained interaction guidance. HyperSkill also incorporates utility-based pruning and structure-informed merging, although it does not enforce a fixed capacity. The paper should better establish what is novel in its architecture and management policies beyond combining existing ideas with capacity limits.
  2. The baselines do not impose capacity constraints. Consequently, the reported storage savings do not establish the advantage of the proposed selection policies. Comparisons against G-Memory with FIFO, LRU, or reservoir sampling, and against simpler eviction policies within FiniteMem, are needed.
  3. Removing a memory layer or disabling its replay shows that the component helps, but not that the particular competition or inhibitory-control formulation is necessary. Hyperparameter sweeps likewise do not establish superiority over simpler alternatives.
  4. Figure 2 closely resembles G-Memory’s overview in layout, color scheme, and visual organization. A more distinct presentation would help clarify FiniteMem’s architectural differences and novel components.
Questions for authors

Can you replace the proposed replay policies with simpler alternatives while keeping FiniteMem’s architecture and capacity unchanged, and report variability across runs?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.