Large language model (LLM) multi-agent systems can scale along two distinct dimensions: by increasing the number of agents and by improving through accumulated experience over time. Although prior work has studied these dimensions separately, their interaction under realistic cost constraints remains unclear. In this paper, we introduce a conceptual scaling view of multi-agent systems that jointly considers team size and lifelong learning ability, and we study how memory design shares this landscape. To this end, we propose LLMA-Mem, a lifelong memory framework for LLM multi-agent systems under flexible memory topologies. We evaluate LLMA-Mem on MultiAgentBench across coding, research, and database environments. Empirically, LLMA-Mem consistently improves long-horizon performance over baselines while reducing cost. Our analysis further reveals a non-monotonic scaling landscape: larger teams do not always produce better long-term performance, and smaller teams can outperform larger ones when memory better supports the reuse of experience. These findings position memory design as a practical path for scaling multi-agent systems more effectively and more efficiently over time.
Figures & tables
Figure 1: Scaling space and cost comparison of LLMA-Mem. Top: LLMA-Mem enabled scaling space for multi-agent systems. Bottom: Average token usage per task, where LLMA-Mem shows substantial cost reduction compared to baselines across models.
Figure 2: CMA curves for Claude-Sonnet-4.5 and Qwen3-next-80B, showing limited long-horizon adaptation.
Figure 3: LLMA-Mem maintains three memory components: episodic memory for task experiences, procedural memory for consolidated reusable strategies, and transactive memory for agent capabilities and team coordination. The right panel illustrates three memory topology configurations that determine how memory is distributed and accessed across agents.
Figure 4: During task execution, the system retrieves relevant memories using a relevance–importance score, updates episodic and transactive statistics after each task, and periodically consolidates episodic experiences into reusable procedural knowledge.
Model
Setting
Coding
Research
Database
TS
CS
AAS
TS
CS
AAS
TS
CS
AAS
Claude-Sonnet-4.5
W.o. Memory
64.43
42.29
53.42
75.12
78.97
76.94
69.39
88.03
77.35
MARBLE
63.05
44.71
53.99
75.17
81.84
78.37
69.70
82.96
75.70
A-Mem
61.57
46.41
53.76
75.69
82.47
77.79
69.70
82.31
75.13
LLMA-Mem
65.00
44.75
54.02 (+0.60)
77.62
74.11
77.77 (+0.83)
70.00
88.27
78.52 (+1.17)
Deepseek-v3.2
W.o. Memory
58.69
43.47
52.26
75.53
58.46
65.94
71.00
82.40
75.55
Table 1: Performance across environments. We report Task Score (TS), Coordination Score (CS), and Average of AS (AAS). Best results within each model and environment are boldfaced. For LLMA-Mem, the AAS cell also shows the change relative to the no-memory baseline.
Figure 5: Cumulative moving average (CMA) curves on representative settings. LLMA-Mem shows more stable long-horizon improvement than MARBLE and A-Mem, with especially margins represented by DeepSeek-v3.2 and Qwen3-32B-Instruct.
Model
Team Size
Avg TS
Avg CS
AAS
Avg Tokens
Claude-Sonnet-4.5
1
77.08
49.58
57.78
53,549
3
76.89
48.78
58.96
139,948
5
77.33
58.00
62.57
279,563
7
78.22
53.11
62.21
344,323
DeepSeek-V3.2
1
72.08
48.13
49.01
41,205
3
72.44
52.00
54.48
116,409
Table 2: Influence of team size on performance. We report average task score (TS), coordination score (CS), Average of AS (AAS), and average token usage per task.
Study
Setting
TS
CS
AAS
Topology ( N=5 )
Local
49.68
46.69
49.43
Shared
46.79
45.96
48.62
Hybrid
46.52
40.32
46.92
Consolidation (Local)
N=2
45.86
42.93
45.07
N=5
49.68
46.69
49.43
N=10
47.72
42.36
46.87
Table 3: Ablation results on the coding environment with Qwen3-32B-Instruct. The upper block compares memory topologies under a fixed consolidation interval N=5 . The lower block compares consolidation intervals under the local topology.
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
Figure 6: An example of procedural memory piece from Qwen3-32B-Instruct on coding envrionment.
Existing large language model (LLM) based multi-agent systems have been shown to scale poorly, and we identify three problems: reading burden that grows with the group, useful information diluted by shared work of unknown quality, and fragility to failed agents. We argue that the problems are largely due to agent-centric designs: the coordination structure is wired into fixed workflows, central coordinators, or peer-to-peer messaging, and recent blackboard-style methods that move the coordination into a shared medium relieve only the last problem. To close the gaps, we build ScaleMAS, where asynchronous agents interact with a shared datastore of linked messages via three operations: publishing messages with optional references to earlier ones that they build on, searching for messages by content, and visiting messages by address. Search returns at most k messages, so an agent's read stays bounded at any group size; results are ranked by relevance and by how often they are cited, which we find to be an effective proxy for quality; and messages persist after their producers leave, so the group carries on when individual agents fail. Across four agentic tasks, ScaleMAS outperforms agent-centric and blackboard-style baselines at lower inference compute, continues to scale up to 128 agents, and tolerates losing agents midway.
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.
Chengxin Yu, Zhaoxin Fan, Faguo Wu +3
Beijing Advanced Innovation Center for Future Blockchain and Privacy Computing, School of Artificial Intelligence, Beihang University · Beijing academy of blockchain and edge computing · National University of Defense Technology +1
Self-evolving multi-agent systems (MAS) have emerged as a promising route to LLM agents that continually improve from experience, with persistent memory at their foundation. However, existing designs almost exclusively adopt a centralized repository shared across agents, incurring communication and coordination overhead, raising privacy concerns, and collapsing agent diversity. We propose DecentMem, a decentralized memory framework in which each agent maintains its own dual-pool memory -- an exploitation pool of consolidated past trajectories and an exploration pool of LLM-generated candidates for unseen contexts. The two pools are reweighted online based on stage-wise feedback from an LLM-as-a-judge. Theoretically, we prove that this design guarantees global reachability of the solution space and achieves O(logT) cumulative regret, matching the stochastic bandit lower bound up to constants. In practice, across three MAS frameworks (AutoGen, DyLAN, AgentNet), three Qwen3 backbones (4B/8B/14B), two Gemma4 backbones (E2B/E4B) and five benchmarks spanning math, code, QA, and embodied tasks, DecentMem improves average accuracy by up to 23.8% over the strongest centralized memory baseline and by up to 52.5% over the no-memory baseline, while reducing token usage by up to 49%.