Pith. sign in

REVIEW 4 major objections 4 minor 2 cited by

When long-term memory frameworks are statistically tied on accuracy, the paper argues the choice reduces to cost minimization, and it shows mem0 dominates Graphiti in a cloud-edge distributed agent testbed.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 10:57 UTC pith:DBQUGGVX

load-bearing objection Useful testbed, broken inference: the paper treats p>0.05 as proof of equivalence and its abstract numbers don't match the body. the 4 major comments →

arxiv 2601.07978 v4 pith:DBQUGGVX submitted 2026-01-12 cs.IR

Cost and Accuracy of Long-Term Memory in Distributed Multi-Agent Systems Based on Large Language Models

Benedict Wolff , Jacopo Bennati This is my paper
classification cs.IR
keywords distributed multi-agent systemslong-term memoryLLM agentsmem0Graphitiknowledge graphscost-accuracy trade-offPareto efficiency
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper tries to establish a decision rule for choosing long-term memory backends in distributed, LLM-based multi-agent systems: if two frameworks are statistically indistinguishable in accuracy, the best balance is simply the cheaper one. To support the rule, it builds an independent, reproducible containerized testbed that simulates a cloud-edge network and runs one LoCoMo long-conversation (19 sessions, 199 questions) through mem0 and Graphiti. The results show accuracy is virtually tied—Graphiti is higher in raw percentages (11.1% vs 7.5% unconstrained; 8.0% vs 6.0% constrained) but two-proportion z-tests cannot reject chance (p-values 0.23 and 0.43)—while cost and resource usage strongly favor mem0, which is faster and cheaper by large margins. A further finding is that adding realistic latency, bandwidth limits, and jitter leaves accuracy essentially unchanged, so the cloud-edge deployment itself is not the bottleneck. The paper matters because it offers a practical, statistically grounded way to compare memory frameworks and suggests that expensive graph-based memory may not pay for itself unless a larger benchmark reveals a real accuracy edge.

Core claim

The paper's central claim is that the balance between cost and accuracy in long-term memory is a dominance relation, not a trade-off. It defines a system as strictly dominant when it has lower total cost of ownership and its accuracy difference relative to a competitor is not statistically significant. On that definition, mem0 Pareto-dominates Graphiti: the raw accuracy gap favors Graphiti, but the Wilson confidence intervals overlap and all four z-tests return p-values above 0.05, so the gap is within random variation; meanwhile mem0 uses 82–87% less loading time, 77–233% less CPU in various phases, orders of magnitude less network bandwidth, and about 40% lower total cost in the unconstrai

What carries the argument

The central mechanism is the statistical Pareto efficiency framework: a system S_A strictly dominates S_B if C(S_A) < C(S_B) (financial dominance) and the accuracy difference A(S_B) − A(S_A) is not statistically significant (statistical equivalence), judged with 95% Wilson confidence intervals and two-proportion z-tests at α = 0.05. This framework is what converts a two-objective problem into a single-objective cost minimization; the paper explicitly rejects ratio-based metrics like cost-per-successful-query and accuracy-efficiency score because they would treat the non-significant accuracy edge as a real gain. The supporting machinery is the testbed: a containerized three-service MAS (coord

Load-bearing premise

The load-bearing premise is that a statistically non-significant accuracy difference (p > 0.05) can be treated as evidence of equal accuracy; with only 199 questions and most answers being 'I don't know', the test has low power, so a real accuracy advantage for Graphiti could easily go undetected.

What would settle it

Run the same testbed on all ten LoCoMo conversations (or a purpose-built benchmark with far more answerable questions) and recompute the accuracy comparison with a sufficiently powered test; if Graphiti's accuracy advantage becomes statistically significant — for instance, a two-proportion z-test with p < 0.05 and a confidence interval excluding zero — the Pareto-dominance conclusion collapses into a genuine cost-accuracy trade-off.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • When accuracy is statistically tied, practitioners can select a memory framework on cost alone; in this workload mem0 is that cheaper choice, with roughly 40% lower total cost and 82–87% faster loading.
  • Edge-cloud network constraints typical of a datacenter-to-edge link do not degrade retrieval accuracy; the measured cost of distribution is a small (4–5%) latency penalty rather than an accuracy loss.
  • Graph-based memory's substantially higher resource consumption (up to thousands of percent more network traffic, over 100% more CPU in some phases) does not produce a measurable accuracy benefit in this test, so its cost premium appears unjustified unless a larger or different benchmark shows a real gap.
  • The statistical Pareto method provides a reusable template for future memory comparisons, while cautioning against naive cost-per-correct-answer metrics that mistake sampling noise for a performance difference.
  • The high rate of 'I don't know' answers (32–66%) shows that long-context retention is still the limiting factor, regardless of which framework is chosen.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • My extension: because the equivalence conclusion rests on a small sample (199 questions, many unanswerable), the dominance claim would be strengthened or reversed by the full LoCoMo benchmark; a sufficiently powered test could reveal a real accuracy gap that turns 'dominance' back into a trade-off.
  • My extension: the paper's own limitation note suggests a plain RAG control; if a simple retrieval baseline matches or beats mem0 at a fraction of the cost, then even the 'winner' here may be over-engineered for this workload.
  • My extension: one could generalize the statistical Pareto rule by defining an 'practical equivalence margin' — decide in advance how large an accuracy gap would justify paying more — making the method less sensitive to underpowered tests.
  • My extension: the testbed's answer grading (string + embedding similarity) may be inflating IDK rates; re-evaluating with human judges or a rubric specifically for LoCoMo's temporal questions could change the accuracy numbers for both frameworks.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper introduces an open-source testbed for evaluating long-term memory frameworks in distributed multi-agent systems and reports an empirical comparison of mem0 (vector-based) and Graphiti (graph-based) on the LoCoMo benchmark under unconstrained and constrained network conditions. The body measures token cost, CPU/RAM/disk/network usage, latency, and accuracy, and finds that mem0 is substantially cheaper and faster, while Graphiti has higher raw accuracy (11.1% vs 7.5% unconstrained; 8.0% vs 6.0% constrained). Because the accuracy differences are not statistically significant at α=0.05, the paper defines 'statistical equivalence' and concludes that mem0 Pareto-dominates Graphiti, so cost minimization determines the optimal choice. The abstract additionally claims that mem0, RAG, and full-context reach 77–81% accuracy, and Graphiti and cognee reach 55–56%, but these results do not appear in the body.

Significance. If the results were valid, the testbed and reproducible open-source code would be a useful contribution to DMAS memory evaluation, and the resource/cost measurements would provide practical guidance. However, the central conclusion rests on treating a non-significant p-value as evidence of equivalence, which is logically unsound. The cost table also contains an implausible anomaly, and the abstract conflicts with the body on core accuracy figures. As a result, the paper's main claims — that mem0 is Pareto-optimal and that cost minimization is therefore warranted — are not supported by the evidence presented. The contribution is currently limited by these load-bearing flaws rather than being a reliable benchmark for practitioners.

major comments (4)
  1. [§4.2, 'Formalizing the balance between cost and accuracy'] The definition of 'Statistical Equivalence' as 'the difference A(Graphiti)-A(mem0) was not statistically significant' is logically invalid. Absence of evidence is not evidence of absence. With N=199 and accuracy rates of 6.0–11.1% (Table 4), the two-proportion z-test has very low power; the 95% CI for the unconstrained accuracy difference is roughly (−0.02, +0.09), so the data are compatible with Graphiti being 9 percentage points more accurate than mem0 — a material gap at these baselines. Therefore the conclusion in §4.3 that mem0 is Pareto-optimal does not follow. An equivalence test with a pre-specified margin (e.g., TOST), a power analysis, or substantially more data is required before such a claim can be made.
  2. [Table 1, financial cost] The reported Graphiti loading cost is internally inconsistent: 4,948.5 × 10^3 tokens cost $0.0027, whereas mem0 with 1,017.4 × 10^3 tokens costs $0.1319. At any plausible LLM API pricing, 4.95 million tokens cannot cost $0.0027. This appears to be a decimal or unit error. Because the TCO and Pareto-dominance conclusions depend directly on cost, this anomaly must be corrected and all downstream cost comparisons recomputed.
  3. [Abstract vs. §3/Table 4] The abstract states that 'mem0, RAG, and full-context reach 77% to 81% accuracy, while Graphiti and cognee reach only 55% to 56%,' and it mentions RAG and cognee baselines, a 4–5% latency penalty, and full-context results. The body reports only mem0 and Graphiti, with accuracy rates of 6.0–11.1% (Table 4). These figures are irreconcilable. Either the abstract describes a different set of experiments or major results are missing from the body. As written, the central accuracy claims of the paper cannot be reproduced or checked from the provided data.
  4. [§3 and §4.4, experimental design] The statistical analysis treats the 199 questions as independent samples, but the experiments appear to be a single run per condition, with all questions drawn from one LoCoMo conversation. No replication across runs, conversations, or random seeds is reported. Given the high IDK rates (32–68%) and small absolute correct counts (12–22), the comparison is underpowered and the conclusions about equivalence are not robust. Additional runs or a hierarchical model are needed to support the claimed generalization.
minor comments (4)
  1. [§3, 'Results and analysis'] The text says Graphiti was '3.6% more accurate' and '2% more accurate,' but the table shows percentage-point differences (11.1−7.5=3.6; 8.0−6.0=2.0). Please use percentage points or relative percentages consistently.
  2. [Table 2, resource usage] The sentence about disk usage at the edge ('88.9% less at the edge') is ambiguous: it is unclear whether this refers to the loading phase or the Q&A phase. Clarify the comparison baseline.
  3. [Abstract, latency claim] The abstract mentions a 'modest latency penalty of 4% to 5% under edge-cloud constraints' for vector-based LTM, but the body does not report a latency table or statistical test for latency. Either add the supporting analysis or remove the claim.
  4. [§2.2, testbed description] The name 'qwen2.5:3b-instruct-q4 K M' has an odd spacing; use the correct model identifier. Also, the sentence 'the coordinator relies entirely on tools that are called by the SLM' is confusingly worded; I suggest rewriting.

Circularity Check

0 steps flagged

No load-bearing circularity: the Pareto-optimal claim is an explicit application of the paper's own dominance definition to independently measured cost and accuracy data; the main flaw is statistical (p>0.05 treated as equivalence), not circular.

full rationale

The paper's central derivation chain is: measure cost and accuracy for mem0 and Graphiti on LoCoMo; observe that mem0 is cheaper and that the accuracy difference is not statistically significant; define 'statistical Pareto efficiency' such that a system dominates another if it is cheaper and the accuracy difference is not statistically significant; conclude mem0 is Pareto-optimal. This conclusion does follow from the stated definition and the measurements, but that is a deductive application of an explicitly disclosed criterion, not a hidden reduction of the output to the input. No parameter is fitted to the target result, no prediction is renamed as a fitted quantity, and no load-bearing self-citation or imported uniqueness theorem is used; the only self-referential element is the newly coined 'statistical pareto efficiency framework,' which is transparently defined in Section 4.2. The genuinely problematic step is the interpretation of 'not statistically significant' as 'statistical equivalence' and 'statistically nullified' (Section 4.2, 'Since the performance differential was statistically nullified, the optimization function collapsed to cost minimization'). With N=199 and IDK rates of 32–68% (Table 4), the two-proportion z-test has low power, and the data are compatible with a materially large accuracy advantage for Graphiti. This is a serious statistical validity concern—absence of evidence treated as evidence of absence—but it is not circularity: the non-significant p-value is an empirical input, not an output derived from the conclusion. The paper itself acknowledges the insufficiency in Section 3 ('the current evidence is insufficient to claim superiority over mem0') and in Section 4.4 recommends future work on the high IDK rates, which further underscores that the dominance conclusion rests on a contestable statistical leap rather than on a definitional identity. The abstract's 77–81% accuracy figures are inconsistent with the body's 6–11% rates, but that reporting inconsistency is a correctness and reproducibility risk, not a circular step. Overall, no significant circularity is present; the paper's conclusion is a definitional application of its stated framework, and the weakness is statistical interpretation, not circular reasoning.

Axiom & Free-Parameter Ledger

3 free parameters · 6 axioms · 0 invented entities

The central comparison relies on external benchmark validity, a questionable equivalence assumption, and chosen network/pricing parameters. No new entities are introduced.

free parameters (3)
  • significance threshold = 0.05
    Used to declare accuracy differences statistically non-significant and thus 'equivalent' in the Pareto framework (Section 4.2, Table 6).
  • network constraint values = 200 ms latency, 8 Mbit/s bandwidth, 50 ms jitter
    Chosen to simulate a datacenter-to-edge link; results may change with other values.
  • AWS Fargate pricing rates = $0.04048/vCPU-hour, $0.004445/GB-hour, $0.000109/GB-hour storage, $0.09/GB egress
    External cost model used to convert resource usage to TCO; a different provider or region changes cost conclusions.
axioms (6)
  • domain assumption LoCoMo benchmark is a valid proxy for long-term conversational memory performance.
    The paper uses LoCoMo to measure knowledge retention and treats accuracy on it as the target metric.
  • standard math Two-proportion z-test and Wilson CIs are appropriate for comparing accuracy rates.
    Used in Section 3 to compare accuracy between frameworks and profiles.
  • ad hoc to paper A non-significant difference (p>0.05) is treated as evidence of statistical equivalence.
    Section 4.2 defines dominance using this 'statistical equivalence' condition; it is a contestable assumption, not a standard equivalence test.
  • domain assumption The first LoCoMo conversation (19 sessions, 199 questions) is representative of DMAS memory workloads.
    The experiment loads only the first conversation, and the authors generalize conclusions from this small sample.
  • domain assumption The testbed's WSL/Docker environment and toxiproxy accurately simulate hybrid cloud-edge network conditions.
    The constrained network profile is applied via toxiproxy on a single machine; this may not capture real distributed deployments.
  • domain assumption OpenAI token costs are accurately computed from usage logs.
    The TCO depends on token usage measurements and official pricing, assumed correct though anomalies appear in Table 1.

pith-pipeline@v1.3.0-alltime-deepseek · 16419 in / 13667 out tokens · 120025 ms · 2026-08-03T10:57:09.633501+00:00 · methodology

0 comments
read the original abstract

Long-term memory (LTM) is fundamental to large language model (LLM)-based agents in the emerging Internet of Agents (IoA), where distributed multi-agent systems (DMAS) span cloud and edge networks. Existing evaluations are typically published by framework providers and focus on token usage and latency, rarely accounting for system-level cost or deployment in DMAS. These gaps are addressed with an independent reproducible testbed that evaluates accuracy, latency, CPU time, peak RAM, disk I/O and network usage in a simulated cloud-edge environment. Three venture capital-funded frameworks spanning vector, graph, and hybrid architectures, namely mem0, Graphiti, and cognee, are compared alongside retrieval-augmented generation (RAG) and full-context baselines on the LoCoMo benchmark under unconstrained and constrained network scenarios. Two clusters emerge: mem0, RAG, and full-context reach 77% to 81% accuracy, while Graphiti and cognee reach only 55% to 56%, a gap driven by retrieval incompleteness rather than reasoning failure. The RAG baseline matches the upper cluster at 8.4 times lower total cost of ownership (TCO) than mem0, and both are the only non-dominated backends on the Pareto frontier. Latency and bandwidth constraints as well as jitter leave retrieval quality unchanged for every backend, while vector-based LTM incurs a modest latency penalty of 4% to 5% under edge-cloud constraints. Compression precision rather than context volume determines LTM accuracy, as full-context forwarding underperforms mem0 despite supplying the entire conversation for each question.

Figures

Figures reproduced from arXiv: 2601.07978 by Benedict Wolff, Jacopo Bennati.

Figure 1
Figure 1. Figure 1: testbed system architecture 3 Results and analysis During the unconstrained experiment, Graphiti required 3.96% more OpenAI tokens than mem0. During the constrained experiment, Graphiti cost 32% more tokens [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: response distribution (Correct, Wrong, IDK) for mem0 and Graphiti under unconstrained and [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: financial cost of computation and token consumption for mem0 and Graphiti. Figures compare [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. DeepRefine: Agent-Compiled Knowledge Refinement via Reinforcement Learning

    cs.CL 2026-05 unverdicted novelty 7.0

    DeepRefine refines agent-compiled knowledge bases via multi-turn abductive diagnosis and RL training with a GBD reward, yielding consistent downstream task gains.

  2. Evaluating Memory Condensation Strategies for Coding Agents in Data-Driven Scientific Discovery

    cs.LG 2026-05 unverdicted novelty 6.0

    Empirical evaluation of eight memory condensation strategies on 480 DiscoveryBench tasks finds no significant impact on hypothesis quality but domain-dependent differences in token efficiency.

Reference graph

Works this paper leans on

94 extracted references · 4 canonical work pages · cited by 2 Pith papers

  1. [1]

    A Survey on the Memory Mechanism of Large Language Model-based Agents,

    Z. Zhang, Q. Dai, X. Bo, C. Ma, R. Li, X. Chen, J. Zhu, Z. Dong, and J.-R. Wen, “A Survey on the Memory Mechanism of Large Language Model-based Agents,”ACM Transactions on Information Systems, vol. 43, no. 6, pp. 1–47, Nov. 2025. doi: 10.1145/3748302. [Online]. Available: https://dl.acm.org/doi/10.1145/3748302

    doi:10.1145/3748302 2025

  2. [2]

    Large Language Model based Multi-Agents: A Survey of Progress and Challenges,

    T. Guo, X. Chen, Y . Wang, R. Chang, S. Pei, N. V . Chawla, O. Wiest, and X. Zhang, “Large Language Model based Multi-Agents: A Survey of Progress and Challenges,” Apr. 2024, arXiv:2402.01680 [cs]. [Online]. Available: http://arxiv.org/abs/2402.01680

    Pith/arXiv arXiv 2024

  3. [3]

    AutoAgent: A Fully-Automated and Zero-Code Framework for LLM Agents,

    J. Tang, T. Fan, and C. Huang, “AutoAgent: A Fully-Automated and Zero-Code Framework for LLM Agents,” Feb. 2025, arXiv:2502.05957 [cs]. [Online]. Available: http://arxiv.org/abs/2502.05957

    arXiv 2025

  4. [4]

    Grid-Agent: An LLM-Powered Multi-Agent System for Power Grid Control,

    Y . Zhang, A. M. Saber, A. Youssef, and D. Kundur, “Grid-Agent: An LLM-Powered Multi-Agent System for Power Grid Control,” Sep. 2025, arXiv:2508.05702 [cs]. [Online]. Available: http://arxiv.org/abs/2508.05702

    Pith/arXiv arXiv 2025

  5. [5]

    AutoML-Agent: A Multi-Agent LLM Framework for Full-Pipeline AutoML,

    P. Trirat, W. Jeong, and S. J. Hwang, “AutoML-Agent: A Multi-Agent LLM Framework for Full-Pipeline AutoML,” Jun. 2025, arXiv:2410.02958 [cs]. [Online]. Available: http: //arxiv.org/abs/2410.02958

    Pith/arXiv arXiv 2025

  6. [6]

    EduMAS: A Novel LLM-Powered Multi-Agent Framework for Educational Support,

    Q. Li, Y . Xie, S. Chakravarty, and D. Lee, “EduMAS: A Novel LLM-Powered Multi-Agent Framework for Educational Support,” in2024 IEEE International Conference on Big Data (BigData). Washington, DC, USA: IEEE, Dec. 2024. doi: 10.1109/BigData62323.2024.10826103. ISBN 979-8-3503-6248-0 pp. 8309–8316. [Online]. Available: https://ieeexplore.ieee.org/document/10826103/

    arXiv 2024

  7. [7]

    LLM4Review: A Multi-Agent Framework for Autonomous Peer Review of AI-Written Research,

    G. Li, “LLM4Review: A Multi-Agent Framework for Autonomous Peer Review of AI-Written Research,” 2025. [Online]. Available: https://openreview.net/forum?id=7dJ7BFv9AT

    2025

  8. [8]

    Moralbench: A Multi-Faceted Benchmark for Ethical and Safety Alignment in Llms,

    M. Gaikwad, “Moralbench: A Multi-Faceted Benchmark for Ethical and Safety Alignment in Llms,”

  9. [9]

    LLM-based multi-agent poetry generation in non-cooperative environments,

    R. Zhang and S. Eger, “LLM-based multi-agent poetry generation in non-cooperative environments,” Sep. 2024, arXiv:2409.03659 [cs]. [Online]. Available: http://arxiv.org/abs/2409.03659

    Pith/arXiv arXiv 2024

  10. [10]

    Balancing performance and cost of LLMs in a multi-agent framework for BIM data retrieval,

    D. Liu, X. Zhou, and Y . Li, “Balancing performance and cost of LLMs in a multi-agent framework for BIM data retrieval,”Architectural Engineering and Design Management, pp. 1–18, Jan. 2025. doi: 10.1080/17452007.2025.2456768. [Online]. Available: https: //www.tandfonline.com/doi/full/10.1080/17452007.2025.2456768

    arXiv 2025

  11. [11]

    TradingAgents: Multi-Agents LLM Financial Trading Framework,

    Y . Xiao, E. Sun, D. Luo, and W. Wang, “TradingAgents: Multi-Agents LLM Financial Trading Framework,” Jun. 2025, arXiv:2412.20138 [q-fin]. [Online]. Available: http://arxiv.org/abs/2412.20138

    Pith/arXiv arXiv 2025

  12. [12]

    LLM-based Multi-Agent Systems: Techniques and Business Perspectives,

    Y . Yang, Q. Peng, J. Wang, Y . Wen, and W. Zhang, “LLM-based Multi-Agent Systems: Techniques and Business Perspectives,” Dec. 2024, arXiv:2411.14033 [cs]. [Online]. Available: http://arxiv.org/abs/2411.14033

    Pith/arXiv arXiv 2024

  13. [13]

    DAWN: Designing Distributed Agents in a Worldwide Network,

    Z. Aminiranjbar, J. Tang, Q. Wang, S. Pant, and M. Viswanathan, “DAWN: Designing Distributed Agents in a Worldwide Network,”IEEE Access, vol. 13, pp. 138 795–138 812, 2025. doi: 10.1109/ACCESS.2025.3588425. [Online]. Available: https://ieeexplore.ieee.org/document/11078243/

    arXiv 2025

  14. [14]

    Evaluating Very Long-Term Conversational Memory of LLM Agents,

    A. Maharana, D.-H. Lee, S. Tulyakov, M. Bansal, F. Barbieri, and Y . Fang, “Evaluating Very Long-Term Conversational Memory of LLM Agents,” Feb. 2024, arXiv:2402.17753 [cs]. [Online]. Available: http://arxiv.org/abs/2402.17753

    Pith/arXiv arXiv 2024

  15. [15]

    MIRIX: Multi-Agent Memory System for LLM-Based Agents,

    Y . Wang and X. Chen, “MIRIX: Multi-Agent Memory System for LLM-Based Agents,” Jul. 2025, arXiv:2507.07957 [cs]. [Online]. Available: http://arxiv.org/abs/2507.07957 16

    Pith/arXiv arXiv 2025

  16. [16]

    Zep: A Temporal Knowledge Graph Architecture for Agent Memory,

    P. Rasmussen, P. Paliychuk, T. Beauvais, J. Ryan, and D. Chalef, “Zep: A Temporal Knowledge Graph Architecture for Agent Memory,” Jan. 2025, arXiv:2501.13956 [cs]. [Online]. Available: http://arxiv.org/abs/2501.13956

    Pith/arXiv arXiv 2025

  17. [17]

    Optimizing the Interface Between Knowledge Graphs and LLMs for Complex Reasoning,

    V . Markovic, L. Obradovic, L. Hajdu, and J. Pavlovic, “Optimizing the Interface Between Knowledge Graphs and LLMs for Complex Reasoning,” May 2025, arXiv:2505.24478 [cs]. [Online]. Available: http://arxiv.org/abs/2505.24478

    Pith/arXiv arXiv 2025

  18. [18]

    Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory,

    P. Chhikara, D. Khant, S. Aryan, T. Singh, and D. Yadav, “Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory,” Apr. 2025, arXiv:2504.19413 [cs]. [Online]. Available: http://arxiv.org/abs/2504.19413

    Pith/arXiv arXiv 2025

  19. [19]

    Minimizing Hallucinations and Communication Costs: Adversarial Debate and V oting Mechanisms in LLM-Based Multi-Agents,

    Y . Yang, Y . Ma, H. Feng, Y . Cheng, and Z. Han, “Minimizing Hallucinations and Communication Costs: Adversarial Debate and V oting Mechanisms in LLM-Based Multi-Agents,”Applied Sciences, vol. 15, no. 7, p. 3676, Mar. 2025. doi: 10.3390/app15073676. [Online]. Available: https://www.mdpi.com/2076-3417/15/7/3676

    work page doi:10.3390/app15073676 2025

  20. [20]

    A Survey of LLM-based Agents: Theories, Technologies, Applications and Suggestions,

    X. Dong, X. Zhang, W. Bu, D. Zhang, and F. Cao, “A Survey of LLM-based Agents: Theories, Technologies, Applications and Suggestions,” in2024 3rd International Conference on Artificial Intelligence, Internet of Things and Cloud Computing Technology (AIoTC). Wuhan, China: IEEE, Sep

  21. [21]

    A survey on LLM-based multi-agent systems: workflow, infrastructure, and challenges,

    X. Li, S. Wang, S. Zeng, Y . Wu, and Y . Yang, “A survey on LLM-based multi-agent systems: workflow, infrastructure, and challenges,”Vicinagearth, vol. 1, no. 1, p. 9, Oct. 2024. doi: 10.1007/s44336-024- 00009-2. [Online]. Available: https://link.springer.com/10.1007/s44336-024-00009-2

    doi:10.1007/s44336-024- 2024

  22. [22]

    A Survey on the Optimization of Large Language Model-based Agents,

    S. Du, J. Zhao, J. Shi, Z. Xie, X. Jiang, Y . Bai, and L. He, “A Survey on the Optimization of Large Language Model-based Agents,” Mar. 2025, arXiv:2503.12434 [cs]. [Online]. Available: http://arxiv.org/abs/2503.12434

    arXiv 2025

  23. [23]

    Optimizing Web-Based AI Query Retrieval with GPT Integration in LangChain A CoT-Enhanced Prompt Engineering Approach,

    W. Guan and Y . Fang, “Optimizing Web-Based AI Query Retrieval with GPT Integration in LangChain A CoT-Enhanced Prompt Engineering Approach,” inProceedings of the 2nd International Conference on Machine Learning and Automation, CONF-MLA 2024, November 21, 2024, Adana, Turkey, 2025. doi: 10.4108/eai.21-11-2024.2354589 ArXiv:2506.15512 [cs]. [Online]. Avail...

    arXiv 2024

  24. [24]

    Cost-Effective Online Multi-LLM Selection with Versatile Reward Models,

    X. Dai, J. Li, X. Liu, A. Yu, and J. C. S. Lui, “Cost-Effective Online Multi-LLM Selection with Versatile Reward Models,” Oct. 2024, arXiv:2405.16587 [cs]. [Online]. Available: http://arxiv.org/abs/2405.16587

    Pith/arXiv arXiv 2024

  25. [25]

    Cost-Efficient Serving of LLM Agents via Test-Time Plan Caching,

    Q. Zhang, M. Wornow, and K. Olukotun, “Cost-Efficient Serving of LLM Agents via Test-Time Plan Caching,” Jun. 2025, arXiv:2506.14852 [cs]. [Online]. Available: http://arxiv.org/abs/2506.14852

    arXiv 2025

  26. [26]

    Efficient Agents: Building Effective Agents While Reducing Cost,

    N. Wang, X. Hu, P. Liu, H. Zhu, Y . Hou, H. Huang, S. Zhang, J. Yang, J. Liu, G. Zhang, C. Zhang, J. Wang, Y . E. Jiang, and W. Zhou, “Efficient Agents: Building Effective Agents While Reducing Cost,” Jul. 2025, arXiv:2508.02694 [cs]. [Online]. Available: http://arxiv.org/abs/2508.02694

    Pith/arXiv arXiv 2025

  27. [27]

    WirelessAgent: Large Language Model Agents for Intelligent Wireless Networks,

    J. Tong, W. Guo, J. Shao, Q. Wu, Z. Li, Z. Lin, and J. Zhang, “WirelessAgent: Large Language Model Agents for Intelligent Wireless Networks,” May 2025, arXiv:2505.01074 [eess]. [Online]. Available: http://arxiv.org/abs/2505.01074

    Pith/arXiv arXiv 2025

  28. [28]

    NetLLM: Adapting Large Language Models for Networking,

    D. Wu, X. Wang, Y . Qiao, Z. Wang, J. Jiang, S. Cui, and F. Wang, “NetLLM: Adapting Large Language Models for Networking,” inProceedings of the ACM SIGCOMM 2024 Conference. Sydney NSW Australia: ACM, Aug. 2024. doi: 10.1145/3651890.3672268. ISBN 979-8-4007-0614-1 pp. 661–678. [Online]. Available: https://dl.acm.org/doi/10.1145/3651890.3672268 17

    arXiv 2024

  29. [29]

    Edge Agentic AI Framework for Autonomous Network Optimisation in O-RAN,

    A. Salama, Z. Nezami, M. M. H. Qazzaz, M. Hafeez, and S. A. R. Zaidi, “Edge Agentic AI Framework for Autonomous Network Optimisation in O-RAN,” Aug. 2025, arXiv:2507.21696 [eess]. [Online]. Available: http://arxiv.org/abs/2507.21696

    Pith/arXiv arXiv 2025

  30. [30]

    Hackphyr: A Local Fine-Tuned LLM Agent for Network Security Environments,

    M. Rigaki, C. Catania, and S. Garcia, “Hackphyr: A Local Fine-Tuned LLM Agent for Network Security Environments,” Sep. 2024, arXiv:2409.11276 [cs]. [Online]. Available: http://arxiv.org/abs/2409.11276

    Pith/arXiv arXiv 2024

  31. [31]

    An Agentic Approach For Dynamic Software-Defined Network Management Using Large Language Models,

    A. S. Araujo, J. M. O. Das Mercˆes, R. L. Da Silva, A. V . De Alencar, I. F. Passos, M. P. Sousa, M. C. Dias, T. F. Meneses, and D. F. S. Santos, “An Agentic Approach For Dynamic Software-Defined Network Management Using Large Language Models,” in2024 IEEE Conference on Network Function Virtualization and Software Defined Networks (NFV-SDN). Natal, Brazil...

    arXiv 2024

  32. [32]

    Model-Distributed Inference for Large Language Models at the Edge,

    D. Macario, H. Seferoglu, and E. Koyuncu, “Model-Distributed Inference for Large Language Models at the Edge,” May 2025, arXiv:2505.18164 [cs]. [Online]. Available: http://arxiv.org/abs/2505.18164

    Pith/arXiv arXiv 2025

  33. [33]

    Distributed Inference of Large Language Models on Edge Devices,

    K. Namboori, R. P. Suresh, S. Hj, S. Mohanty, and J. Rangareddy, “Distributed Inference of Large Language Models on Edge Devices,” inProceedings of the 2025 14th International Conference on Software and Computer Applications. Kuala Lumpur Malaysia: ACM, Feb

    2025

  34. [34]

    Adaptive AI Agent Placement and Migration in Edge Intelligence Systems,

    X. Wang, J. He, Z. Tang, J. Guo, J. Lou, L. Qian, T. Wang, and W. Jia, “Adaptive AI Agent Placement and Migration in Edge Intelligence Systems,” Aug. 2025, arXiv:2508.03345 [cs]. [Online]. Available: http://arxiv.org/abs/2508.03345

    Pith/arXiv arXiv 2025

  35. [35]

    Vector Storage Based Long-term Memory Research on LLM,

    K. Li, X. Jing, and C. Jing, “Vector Storage Based Long-term Memory Research on LLM,”International Journal of Advanced Network, Monitoring and Controls, vol. 9, no. 3, pp. 69–79, Sep. 2024. doi: 10.2478/ijanmc-2024-0029. [Online]. Available: https: //www.sciendo.com/article/10.2478/ijanmc-2024-0029

    work page doi:10.2478/ijanmc-2024-0029 2024

  36. [36]

    ISBN 979-8-4007-1012-4 pp

    doi: 10.1145/3731806.3731859. ISBN 979-8-4007-1012-4 pp. 321–326. [Online]. Available: https://dl.acm.org/doi/10.1145/3731806.3731859

    arXiv

  37. [37]

    LLMs Working in Harmony: A Survey on the Technological Aspects of Building Effective LLM-Based Multi Agent Systems,

    R. M. Aratchige and W. M. K. S. Ilmini, “LLMs Working in Harmony: A Survey on the Technological Aspects of Building Effective LLM-Based Multi Agent Systems,” Mar. 2025, arXiv:2504.01963 [cs]. [Online]. Available: http://arxiv.org/abs/2504.01963

    Pith/arXiv arXiv 2025

  38. [38]

    LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory,

    D. Wu, H. Wang, W. Yu, Y . Zhang, K.-W. Chang, and D. Yu, “LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory,” 2024, version Number: 2. [Online]. Available: https://arxiv.org/abs/2410.10813

    Pith/arXiv arXiv 2024

  39. [39]

    HiAgent: Hierarchical Working Memory Management for Solving Long-Horizon Agent Tasks with Large Language Model,

    M. Hu, T. Chen, Q. Chen, Y . Mu, W. Shao, and P. Luo, “HiAgent: Hierarchical Working Memory Management for Solving Long-Horizon Agent Tasks with Large Language Model,” Aug. 2024, arXiv:2408.09559 [cs]. [Online]. Available: http://arxiv.org/abs/2408.09559

    Pith/arXiv arXiv 2024

  40. [40]

    StoryBench: A Dynamic Benchmark for Evaluating Long-Term Memory with Multi Turns,

    L. Wan and W. Ma, “StoryBench: A Dynamic Benchmark for Evaluating Long-Term Memory with Multi Turns,” 2025, version Number: 1. [Online]. Available: https://arxiv.org/abs/2506.13356

    Pith/arXiv arXiv 2025

  41. [41]

    Multi-Agent Collaboration Mechanisms: A Survey of LLMs,

    K.-T. Tran, D. Dao, M.-D. Nguyen, Q.-V . Pham, B. O’Sullivan, and H. D. Nguyen, “Multi-Agent Collaboration Mechanisms: A Survey of LLMs,” Jan. 2025, arXiv:2501.06322 [cs]. [Online]. Available: http://arxiv.org/abs/2501.06322

    Pith/arXiv arXiv 2025

  42. [42]

    LongBench v2: Towards Deeper Understanding and Reasoning on Realistic Long-context Multitasks,

    Y . Bai, S. Tu, J. Zhang, H. Peng, X. Wang, X. Lv, S. Cao, J. Xu, L. Hou, Y . Dong, J. Tang, and J. Li, “LongBench v2: Towards Deeper Understanding and Reasoning on Realistic Long-context Multitasks,” Jan. 2025, arXiv:2412.15204 [cs]. [Online]. Available: http://arxiv.org/abs/2412.15204

    Pith/arXiv arXiv 2025

  43. [43]

    AgentsNet: Coordination and Collaborative Reasoning in Multi-Agent LLMs,

    F. Gr¨otschla, L. M ¨uller, J. T ¨onshoff, M. Galkin, and B. Perozzi, “AgentsNet: Coordination and Collaborative Reasoning in Multi-Agent LLMs,” Jul. 2025, arXiv:2507.08616 [cs]. [Online]. Available: http://arxiv.org/abs/2507.08616

    Pith/arXiv arXiv 2025

  44. [44]

    A Dynamic LLM-Powered Agent Network for Task-Oriented Agent Collaboration,

    Z. Liu, Y . Zhang, P. Li, Y . Liu, and D. Yang, “A Dynamic LLM-Powered Agent Network for Task-Oriented Agent Collaboration,” Nov. 2024, arXiv:2310.02170 [cs]. [Online]. Available: http://arxiv.org/abs/2310.02170

    Pith/arXiv arXiv 2024

  45. [45]

    AgentNet: Decentralized Evolutionary Coordination for LLM-based Multi-Agent Systems,

    Y . Yang, H. Chai, S. Shao, Y . Song, S. Qi, R. Rui, and W. Zhang, “AgentNet: Decentralized Evolutionary Coordination for LLM-based Multi-Agent Systems,” May 2025, arXiv:2504.00587 [cs]. [Online]. Available: http://arxiv.org/abs/2504.00587 18

    Pith/arXiv arXiv 2025

  46. [46]

    BudgetMLAgent: A Cost-Effective LLM Multi-Agent system for Automating Machine Learning Tasks,

    S. Gandhi, M. Patwardhan, L. Vig, and G. Shroff, “BudgetMLAgent: A Cost-Effective LLM Multi-Agent system for Automating Machine Learning Tasks,” Jan. 2025, arXiv:2411.07464 [cs]. [Online]. Available: http://arxiv.org/abs/2411.07464

    Pith/arXiv arXiv 2025

  47. [47]

    DynTaskMAS: A Dynamic Task Graph-driven Framework for Asynchronous and Parallel LLM-based Multi-Agent Systems,

    J. Yu, Y . Ding, and H. Sato, “DynTaskMAS: A Dynamic Task Graph-driven Framework for Asynchronous and Parallel LLM-based Multi-Agent Systems,” Mar. 2025, arXiv:2503.07675 [cs]. [Online]. Available: http://arxiv.org/abs/2503.07675

    arXiv 2025

  48. [48]

    AgentTaxo: Dissecting and Benchmarking Token Distribution of LLM Multi-Agent Systems,

    Q. Wang, Z. Tang, Z. JIANG, N. Chen, T. Wang, and B. He, “AgentTaxo: Dissecting and Benchmarking Token Distribution of LLM Multi-Agent Systems,” inICLR 2025 Workshop on Foundation Models in the Wild, 2025. [Online]. Available: https://openreview.net/forum?id=0iLbiYYIpC

    2025

  49. [49]

    LLM-Based Multi-Agent Decision-Making: Challenges and Future Directions,

    C. Sun, S. Huang, and D. Pompili, “LLM-Based Multi-Agent Decision-Making: Challenges and Future Directions,”IEEE Robotics and Automation Letters, vol. 10, no. 6, pp. 5681–5688, Jun. 2025. doi: 10.1109/LRA.2025.3562371. [Online]. Available: https://ieeexplore.ieee.org/document/10970024/

    arXiv 2025

  50. [50]

    Why Do Multi-Agent LLM Systems Fail?

    M. Cemri, M. Z. Pan, S. Yang, L. A. Agrawal, B. Chopra, R. Tiwari, K. Keutzer, A. Parameswaran, D. Klein, K. Ramchandran, M. Zaharia, J. E. Gonzalez, and I. Stoica, “Why Do Multi-Agent LLM Systems Fail?” Apr. 2025, arXiv:2503.13657 [cs]. [Online]. Available: http://arxiv.org/abs/2503.13657

    Pith/arXiv arXiv 2025

  51. [51]

    G-Designer: Architecting Multi-agent Communication Topologies via Graph Neural Networks,

    G. Zhang, Y . Yue, X. Sun, G. Wan, M. Yu, J. Fang, K. Wang, T. Chen, and D. Cheng, “G-Designer: Architecting Multi-agent Communication Topologies via Graph Neural Networks,” Feb. 2025, arXiv:2410.11782 [cs]. [Online]. Available: http://arxiv.org/abs/2410.11782

    Pith/arXiv arXiv 2025

  52. [52]

    NetSafe: Exploring the Topological Safety of Multi-agent Networks,

    M. Yu, S. Wang, G. Zhang, J. Mao, C. Yin, Q. Liu, Q. Wen, K. Wang, and Y . Wang, “NetSafe: Exploring the Topological Safety of Multi-agent Networks,” Oct. 2024, arXiv:2410.15686 [cs]. [Online]. Available: http://arxiv.org/abs/2410.15686

    Pith/arXiv arXiv 2024

  53. [53]

    Autonomous network monitoring using LLMs and multi-agent systems,

    Oluwatosin Oladayo Aramide, “Autonomous network monitoring using LLMs and multi-agent systems,” World Journal of Advanced Engineering Technology and Sciences, vol. 13, no. 2, pp. 974–985, Dec

  54. [54]

    G-Safeguard: A Topology-Guided Security Lens and Treatment on LLM-based Multi-agent Systems,

    S. Wang, G. Zhang, M. Yu, G. Wan, F. Meng, C. Guo, K. Wang, and Y . Wang, “G-Safeguard: A Topology-Guided Security Lens and Treatment on LLM-based Multi-agent Systems,” Feb. 2025, arXiv:2502.11127 [cs]. [Online]. Available: http://arxiv.org/abs/2502.11127

    Pith/arXiv arXiv 2025

  55. [55]

    Maestro: LLM-Driven Collaborative Automation of Intent-Based 6G Networks,

    I. Chatzistefanidis, A. Leone, and N. Nikaein, “Maestro: LLM-Driven Collaborative Automation of Intent-Based 6G Networks,”IEEE Networking Letters, vol. 6, no. 4, pp. 227–231, Dec. 2024. doi: 10.1109/LNET.2024.3503292. [Online]. Available: https://ieeexplore.ieee.org/document/10758700/

    arXiv 2024

  56. [56]

    Network Function Orchestration with LLM based Multi-Agent System,

    J. Gemayel and A. Mokh, “Network Function Orchestration with LLM based Multi-Agent System,” in2025 IEEE International Conference on Communications Workshops (ICC Workshops). Montreal, QC, Canada: IEEE, Jun. 2025. doi: 10.1109/ICCWorkshops67674.2025.11162489. ISBN 979-8-3315-9624-8 pp. 262–267. [Online]. Available: https://ieeexplore.ieee.org/document/11162489/ 19

    arXiv 2025

  57. [57]

    [Online]

    doi: 10.30574/wjaets.2024.13.2.0639. [Online]. Available: https://wjaets.com/node/3518

    work page doi:10.30574/wjaets.2024.13.2.0639 2024

  58. [58]

    GenOnet: Generative Open xG Network Simulation with Multi-Agent LLM and ns-3,

    F. Rezazadeh, A. A. Gargari, S. Lag ´en, J. Mangues-Bafalluy, D. Niyato, and L. Liu, “GenOnet: Generative Open xG Network Simulation with Multi-Agent LLM and ns-3,” in2024 3rd International Conference on 6G Networking (6GNet). Paris, France: IEEE, Oct. 2024. doi: 10.1109/6GNet63182.2024.10765766. ISBN 979-8-3503-7859-7 pp. 69–71. [Online]. Available: http...

    arXiv 2024

  59. [59]

    Distributed Multi-agent Systems Based on the Mixture of Experts Architecture in the Context of 6G Wireless Technologies,

    V . Slyusar, “Distributed Multi-agent Systems Based on the Mixture of Experts Architecture in the Context of 6G Wireless Technologies,” inApplied Innovations in Information and Communication Technology, S. Dovgyi, E. Siemens, L. Globa, O. Kopiika, and O. Stryzhak, Eds. Cham: Springer Nature Switzerland, 2025, vol. 1338, pp. 81–110. ISBN 978-3-031-89295-0 ...

    work page doi:10.1007/978-3-031-89296-7 2025

  60. [60]

    Multi-Agent Reinforcement Learning in Wireless Distributed Networks for 6G,

    J. Zhang, Z. Liu, Y . Zhu, E. Shi, B. Xu, C. Yuen, D. Niyato, M. Debbah, S. Jin, B. Ai, Xuemin, and Shen, “Multi-Agent Reinforcement Learning in Wireless Distributed Networks for 6G,” Feb. 2025, arXiv:2502.05812 [cs]. [Online]. Available: http://arxiv.org/abs/2502.05812

    Pith/arXiv arXiv 2025

  61. [61]

    AI-Driven Decentralized Network Management: Leveraging Multi-Agent Large Language Models for Scalable Optimization,

    H. Lee, M. Kim, S. Baek, W. Zhou, M. Debbah, and I. Lee, “AI-Driven Decentralized Network Management: Leveraging Multi-Agent Large Language Models for Scalable Optimization,”IEEE Communications Magazine, vol. 63, no. 6, pp. 50–56, Jun. 2025. doi: 10.1109/MCOM.001.2400577. [Online]. Available: https://ieeexplore.ieee.org/document/11018287/

    arXiv 2025

  62. [62]

    LLM-Driven Multi-Agent Architectures for Intelligent Self-Organizing Networks,

    A. Qayyum, A. Albaseer, J. Qadir, A. Al-Fuqaha, and M. Abdallah, “LLM-Driven Multi-Agent Architectures for Intelligent Self-Organizing Networks,”IEEE Network, pp. 1–10, 2025. doi: 10.1109/MNET.2025.3605319. [Online]. Available: https://ieeexplore.ieee.org/document/11169757/

    arXiv 2025

  63. [63]

    G-Memory: Tracing Hierarchical Memory for Multi-Agent Systems,

    G. Zhang, M. Fu, G. Wan, M. Yu, K. Wang, and S. Yan, “G-Memory: Tracing Hierarchical Memory for Multi-Agent Systems,” 2025, version Number: 2. [Online]. Available: https://arxiv.org/abs/2506.07398

    Pith/arXiv arXiv 2025

  64. [64]

    AGENTiGraph: A Multi-Agent Knowledge Graph Framework for Interactive, Domain-Specific LLM Chatbots,

    X. Zhao, M. Blum, F. Gao, Y . Chen, B. Yang, L. Marquez-Carpintero, M. Pina-Navarro, Y . Fu, S. Morikawa, Y . Iwasawa, Y . Matsuo, C. Park, and I. Li, “AGENTiGraph: A Multi-Agent Knowledge Graph Framework for Interactive, Domain-Specific LLM Chatbots,” Aug. 2025, arXiv:2508.02999 [cs]. [Online]. Available: http://arxiv.org/abs/2508.02999

    Pith/arXiv arXiv 2025

  65. [65]

    Large Language Model Enhanced Multi-Agent Systems for 6G Communications,

    F. Jiang, Y . Peng, L. Dong, K. Wang, K. Yang, C. Pan, D. Niyato, and O. A. Dobre, “Large Language Model Enhanced Multi-Agent Systems for 6G Communications,”IEEE Wireless Communications, vol. 31, no. 6, pp. 48–55, Dec. 2024. doi: 10.1109/MWC.016.2300600. [Online]. Available: https://ieeexplore.ieee.org/document/10638533/

    arXiv 2024

  66. [66]

    LLM Multi-Agent Systems: Challenges and Open Problems,

    S. Han, Q. Zhang, Y . Yao, W. Jin, and Z. Xu, “LLM Multi-Agent Systems: Challenges and Open Problems,” May 2025, arXiv:2402.03578 [cs]. [Online]. Available: http://arxiv.org/abs/2402.03578

    Pith/arXiv arXiv 2025

  67. [67]

    Context Adaptive Memory-Efficient LLM Inference for Edge Multi-Agent Systems,

    H. Mohammed, H. Yin, and S. C. Boyapati, “Context Adaptive Memory-Efficient LLM Inference for Edge Multi-Agent Systems,” inProceedings of the 24th International Conference on Autonomous Agents and Multiagent Systems, ser. AAMAS ’25. Richland, SC: International Foundation for Autonomous Agents and Multiagent Systems, 2025. ISBN 979-8-4007-1426-9 pp. 2678–2...

    2025

  68. [68]

    Achilles Heel of Distributed Multi-Agent Systems,

    Y . Zhang, Y . Li, T. Zhao, K. Zhu, H. Wang, and N. Vasconcelos, “Achilles Heel of Distributed Multi-Agent Systems,” Apr. 2025, arXiv:2504.07461 [cs]. [Online]. Available: http://arxiv.org/abs/2504.07461

    Pith/arXiv arXiv 2025

  69. [69]

    Intrinsic Memory Agents: Heterogeneous Multi-Agent LLM Systems through Structured Contextual Memory,

    S. Yuen, F. G. Medina, T. Su, Y . Du, and A. J. Sobey, “Intrinsic Memory Agents: Heterogeneous Multi-Agent LLM Systems through Structured Contextual Memory,” Aug. 2025, arXiv:2508.08997 [cs]. [Online]. Available: http://arxiv.org/abs/2508.08997

    arXiv 2025

  70. [70]

    AutoHMA-LLM: Efficient Task Coordination and Execution in Heterogeneous Multi-Agent Systems Using Hybrid Large Language Models,

    T. Yang, P. Feng, Q. Guo, J. Zhang, X. Zhang, J. Ning, X. Wang, and Z. Mao, “AutoHMA-LLM: Efficient Task Coordination and Execution in Heterogeneous Multi-Agent Systems Using Hybrid Large Language Models,”IEEE Transactions on Cognitive Communications and Networking, vol. 11, no. 2, pp. 987–998, Apr. 2025. doi: 10.1109/TCCN.2025.3528892. [Online]. Availabl...

    arXiv 2025

  71. [71]

    AgentRAN: An Agentic AI Architecture for Autonomous Control of Open 6G Networks,

    M. Elkael, S. D’Oro, L. Bonati, M. Polese, Y . Lee, K. Furueda, and T. Melodia, “AgentRAN: An Agentic AI Architecture for Autonomous Control of Open 6G Networks,” Aug. 2025, arXiv:2508.17778 [cs]. [Online]. Available: http://arxiv.org/abs/2508.17778

    arXiv 2025

  72. [72]

    Bock and B

    P. Bock and B. Scheibe,Getting it right: R&D methods for science and engineering. San Diego: Academic Press, 2001. ISBN 978-0-12-108852-1

    2001

  73. [73]

    Secure Multi-LLM Agentic AI and Agentification for Edge General 20 Intelligence by Zero-Trust: A Survey,

    Y . Liu, R. Zhang, H. Luo, Y . Lin, G. Sun, D. Niyato, H. Du, Z. Xiong, Y . Wen, A. Jamalipour, D. I. Kim, and P. Zhang, “Secure Multi-LLM Agentic AI and Agentification for Edge General 20 Intelligence by Zero-Trust: A Survey,” Aug. 2025, arXiv:2508.19870 [cs]. [Online]. Available: http://arxiv.org/abs/2508.19870

    Pith/arXiv arXiv 2025

  74. [74]

    LLM Enabled Multi-Agent System for 6G Networks: Framework and Method of Dual-Loop Edge-Terminal Collaboration,

    Z. Qu, W. Wang, Z. Yu, B. Sun, Y . Li, and X. Zhang, “LLM Enabled Multi-Agent System for 6G Networks: Framework and Method of Dual-Loop Edge-Terminal Collaboration,” Sep. 2025, arXiv:2509.04993 [cs]. [Online]. Available: http://arxiv.org/abs/2509.04993

    Pith/arXiv arXiv 2025

  75. [75]

    sentence-transformers,

    Hugging Face, “sentence-transformers,” Nov. 2025. [Online]. Available: https://github.com/ huggingface/sentence-transformers/tree/main

    2025

  76. [76]

    all-MiniLM-L6-v2,

    ——, “all-MiniLM-L6-v2,” Nov. 2025. [Online]. Available: https://huggingface.co/ sentence-transformers/all-MiniLM-L6-v2

    2025

  77. [77]

    Mem0 raises $24M to build the memory layer for AI,

    T. Singh, “Mem0 raises $24M to build the memory layer for AI,” Nov. 2025. [Online]. Available: https://mem0.ai/series-a

    2025

  78. [78]

    Neo4j Announces $325 Million Series F Investment, the Largest in Database History,

    Neo4j, Inc., “Neo4j Announces $325 Million Series F Investment, the Largest in Database History,” Nov. 2025. [Online]. Available: https://neo4j.com/press-releases/neo4j-announces-seriesf-funding/

    2025

  79. [79]

    Quick Start,

    zep, “Quick Start,” Nov. 2025. [Online]. Available: https://help.getzep.com/graphiti/getting-started/ quick-start

    2025

  80. [80]

    qwen2.5:3b-instruct-q4 k m,

    Ollama, “qwen2.5:3b-instruct-q4 k m,” Nov. 2025. [Online]. Available: https://ollama.com/library/ qwen2.5:3b-instruct-q4 K M

    2025

Showing first 80 references.