PaperlayerView exact revision on arXiv

Source-linked AI summary

CoMem: Collective-Individual Memory Synergy for Evolutionary Multi-Agent Systems

Chengxin Yu, Zhaoxin Fan, Faguo Wu, Hongwei Zheng, Yun Zhou, Zhiyu Li

arXiv:2609.15009v1cs.AI

TL;DR

Flat shared memories can accumulate noisy records and erase agent specialization. CoMem combines private experience, curated group knowledge, and dual-stream retrieval; experiments report consistent gains on ALFWorld and PDDL while mitigating memory pollution.

  • Problem

    Many MAS memory approaches use flat structures without explicitly modeling group memory or synergy between collective and individual memories, leaving shared-memory coordination and quality control insufficiently addressed.

    Source: §1 · p. 1

  • Method

    CoMem combines agent-specific validated experience, a collectively curated pool of cross-agent-proven insights, and parallel retrieval from both memory streams.

    Source: §1 · p. 2, §3.1 · p. 3, §3.3 · p. 4

  • Results

    Across frameworks, CoMem consistently improves strong baselines on ALFWorld and PDDL; on MacNet, ALFWorld success rates rise from 79.85 to 89.55.

    Source: §1 · p. 2, §4.2 · p. 6

  • Takeaways & Limitations

    CoMem is reported as a robust, effective approach to mitigating memory pollution while preserving individual learning and sharing validated insights.

    Source: §5 · p. 7

Abstract

from arXiv · show

Designing effective memory mechanisms is crucial for advancing LLM-driven Multi-Agent Systems (MAS), helping agents learn together and perform better over time. While recent work has led to strong cooperation skills, most methods still use flat, unstructured memories, which easily get filled with noise and erase differences between agents. To address this, we introduce the concept of collective-individual memory synergy and propose CoMem, an architecture that unifies both private experience and shared knowledge for multi-agent learning. CoMem features:(i) Private Experience Sedimentation, which lets each agent keep and update its own useful memories over time;(ii) Collective Wisdom Curation, which carefully selects only widely proven ideas to be shared among agents;(iii)Parallel Dual-Stream Retrieval, which allows agents to draw both from their own memory and the group's wisdom, using clustering to ensure diversity.Experiments on ALFWorld and PDDL benchmarks show that CoMem achieves strong overall performance and robustly avoids memory pollution.

1 Introduction

Flat shared memories can accumulate noise and erase agents’ role-specific characteristics. The introduction proposes balancing private learning with screened collective knowledge, then presents CoMem and reports gains on ALFWorld and PDDL.

1 Introduction

  • Flat memory designs risk noise accumulation and behavioral homogenization, which reinforce each other and can worsen performance.Noise comes from unverified or low-quality behaviors; homogenization erodes agents’ individual specializations through repeated cross-writing.Source: Figure 1 · p. 1, §1 · p. 2
  • The proposed group-memory principle balances robust individual learning with meaningful shared knowledge, screening insights before they become collective wisdom.Agents learn, test, and refine experiences in isolation, while consensus is distilled through selective sharing and rigorous screening.Source: §1 · p. 2
  • CoMem connects isolated, agent-specific experience with a curated shared pool through private-memory pruning, cross-agent validation, and retrieval from both memory streams.The introduction describes private memories as retaining frequently validated knowledge and collective memory as admitting insights after repeated cross-agent verification.Source: §1 · p. 2
  • CoMem consistently improves over strong baselines on ALFWorld and PDDL while mitigating memory pollution.Its three mechanisms combine private experience sedimentation, collective wisdom curation, and parallel dual-stream retrieval.Source: §1 · p. 2

2 Related Work

The related work describes capable collaborative MAS frameworks but notes that they often do not preserve shared strategic knowledge across tasks. Existing memory approaches face cross-agent interference, dilution, contamination, and homogenization.

2.1 LLM-based Multi-Agent Systems

  • MAS frameworks support customizable communication or role-specific workflows, but largely discard interaction histories and lack mechanisms for preserving shared strategic knowledge across tasks.The section cites AutoGen and CAMEL for communication topologies, and ChatDev and MetaGPT for role-specific workflows.Source: §2.1 · p. 2

2.2 Memory Mechanisms for Autonomous Agents

  • Single-agent memory supports lifelong learning, but shared MAS designs risk dilution, contamination, and homogenization; G-Memory retains flat cross-agent read/write.The paper positions CoMem’s group-memory approach as addressing this limitation.Source: §2.2 · pp. 2–3

3 Methodology

CoMem addresses pollution and specialization loss in flat shared-memory MAS with separate private and curated collective tiers. It keeps private histories agent-specific while promoting and reviewing knowledge through empirical criteria.

3.1 Overview

  • CoMem addresses flat-memory pollution from unverified behavior and erased specialization by separating private stores from a validated shared pool.Across sequential collaborative tasks, entries move from private memory into the pool, and collective entries that lose utility are purged.Source: §3.1 · p. 3

3.2 Private Experience Sedimentation

  • Agents distill their own episode traces into actionable insights and prune nonpromoted entries after Kmax consecutive misses, preserving each agent’s separate history.The private store filters access by agent identity, preventing cross-agent contamination.Source: §3.2 · p. 4

3.3 Collective Wisdom Curation

  • Promotion requires sufficient successful retrievals and above-baseline average scores; collective entries receive ongoing utility review and persistently weak entries are removed.Utility combines normalized episodic reward with agents’ subjective LLM-assessed insight quality; the curated pool is refined through empirical feedback.Source: §3.3 · p. 4

ALFWorld PDDL ALFWorld PDDL ALFWorld PDDL ALFWorld PDDL

CoMem reports positive No-memory deltas across the listed ALFWorld/PDDL framework configurations. Its retrieval mechanism combines private and collective knowledge in parallel, filtering redundant shared entries for a diverse context.

ALFWorld PDDL ALFWorld PDDL ALFWorld PDDL ALFWorld PDDL

  • CoMem success rates (with No-memory deltas) are 88.81+9.71, 74.44+4.92, 87.31+6.71, 69.50+1.72, 89.55+9.70, 70.19+9.41, 90.30+5.97, and 72.74+2.73.The table reports ALFWorld and PDDL success rates across MAS frameworks; superscripts are absolute percentage-point differences from No-memory.Source: Table 1 · p. 5, ALFWorld PDDL ALFWorld PDDL ALFWorld… · p. 5

3.4 Parallel Dual-Stream Retrieval

  • Agents retrieve private and collective entries in parallel; semantic similarity finds personal task-relevant guidance, while hierarchical clustering filters overlapping shared entries into a succinct, diverse context.The two streams provide individualized know-how and empirically validated group wisdom without sequential lookups.Source: §3.4 · p. 5

4 Experiments

CoMem is evaluated across two benchmarks and four MAS frameworks, where it improves over No-memory; its results also examine token cost and sensitivity to retrieval and memory settings.

4.1 Experiment Setup

  • Evaluations cover ALFWorld and PDDL across AutoGen, DyLAN, MacNet, and CARD, with memory modules added without altering host coordination.Supplementary Qwen3-32B tests also show consistent improvements, supporting generalization across backbone families.Source: §4.1 · pp. 5–6

4.3 Cost Analysis (RQ2)

  • CoMem achieves the highest success rates with modestly increased token consumption, while cooperative baselines can incur comparable or higher costs without consistent gains.The reported average absolute improvements are 8.02 on ALFWorld and 4.70 on PDDL.Source: Figure 3 · p. 6, §4.3 · p. 6

4.2 Main Results (RQ1)

  • 8.02 percentage-point average gain on ALFWorld and 4.70 on PDDL are reported over No-memory across frameworks.On MacNet, success rates rise from 79.85 to 89.55 on ALFWorld and from 60.78 to 70.19 on PDDL; CoMem also improves over the strong CARD baseline.Source: §4.2 · p. 6

4.4 Framework Analysis (RQ3)

  • On AutoGen PDDL, k=5 yields 80.40 versus 74.80 at k=1 (5.60 points); moderate k ∈{3, 5} is robust, but optima vary by task and framework.Larger k injects noise, whereas smaller k misses context.Source: Figure 4 · p. 6, §4.4 · pp. 6–7
  • On AutoGen PDDL, full CoMem scores 74.44 versus 65.10 collective-only and 62.41 personal-only, with removing the collective layer causing the larger drop.The drops are 9.34 points without the private layer and 12.03 without the collective layer; the authors describe the private layer as a slightly stronger safeguard against pollution.Source: Table 2 · p. 7, §4.4 · p. 7
  • On AutoGen, Kmax=5 yields 74.44 on PDDL and Kmax=6 yields 88.81 on ALFWorld; overly small windows over-prune, while overly large ones accumulate noise.The authors report an inverted-U pattern on both benchmarks.Source: Table 3 · p. 7, §4.4 · p. 7
  • At α=0.5, AutoGen reaches 74.44 on PDDL and 88.81 on ALFWorld; too-low α lags environmental change, while too-high α promotes transient noise.Both benchmarks show an inverted-U trend, with α=0.5 optimal among tested settings.Source: Table 3 · p. 7, §4.4 · p. 7

5 Conclusion

CoMem separates agents’ private experiences from a validated shared pool and retrieves from both. Experiments support it as an effective approach to mitigating memory pollution in LLM-based multi-agent systems.

5 Conclusion

  • CoMem separates private agent experience from a validated collective pool and uses parallel dual-stream retrieval to draw on both for collaboration.The paper reports comprehensive experiments demonstrating a robust and effective approach to mitigating memory pollution in LLM-based multi-agent systems.Source: §5 · p. 7

Related papers

Loading 2609.15009v1…