Abstract
Multi-agent systems built on large language models (LLMs) frequently exchange contextual memory, creating potential vectors for poisoning through malicious prompts or corrupted state propagation. This study proposes a self-healing memory architecture combining semantic anomaly detection and consensus-based repair. Memory embeddings are continuously monitored using cosine similarity drift detection relative to historical stable clusters. When abnormal divergence exceeds a threshold (δ = 0.35), a cross-agent consensus mechanism reconstructs corrupted segments through majority voting and redundancy checks. Evaluation was conducted on 18 collaborative reasoning tasks involving 10–20 LLM agents. Under simulated adversarial prompt injection, task accuracy declined by 29.5% in baseline settings but only 8.3% under the proposed repair framework. Memory corruption persistence time decreased from 14.2 interaction rounds to 4.7 rounds. The architecture enhances resilience against semantic memory poisoning in collaborative LLM ecosystems.
References
Chen, H., Li, J., Ma, X., & Mao, Y. (2025, June). Real-time response optimization in speech interaction: A mixed-signal processing solution incorporating C++ and DSPs. In 2025 7th International Conference on Artificial Intelligence Technologies and Applications (ICAITA) (pp. 110-114). IEEE.
Krishnan, N. (2025). Advancing multi-agent systems through model context protocol: Architecture, implementation, and applications. arXiv preprint arXiv:2504.21030.
Liu, H., Xu, D., Ma, Q., Xu, S., & Qiu, D. (2026). Memory Poisoning Propagation and Repair Mechanism in Multi-Agent Collaborative Environments.
Rezazadeh, A., Li, Z., Lou, A., Zhao, Y., Wei, W., & Bao, Y. (2025). Collaborative memory: Multi-user memory sharing in llm agents with dynamic access control. arXiv preprint arXiv:2505.18279.
Li, T., Liu, S., Hong, E., & Xia, J. (2025). Human Resource Optimization in the Hospitality Industry Big Data Forecasting and Cross-Cultural Engagement.
Srivastava, S. S., & He, H. (2025). MemoryGraft: Persistent compromise of LLM agents via poisoned experience retrieval. arXiv preprint arXiv:2512.16962.
Ma, Q., Yue, L., Xu, S., Shi, Y., & Liu, H. (2026). Web Agent Agentic Reinforcement Learning Decision Model Under Multi-Cost and Failure Risk Constraints.
Baroni, L. A., & Pereira, R. (2024, October). Deceptive patterns under a sociotechnical view. In Proceedings of the XXIII Brazilian Symposium on Human Factors in Computing Systems (pp. 1-13).
Qiu, D., Xu, D., & Yue, L. (2025, December). Reinforcement Learning-Augmented LLM Agents for Collaborative Decision Making and Performance Optimization. In 2025 7th International Conference on Frontier Technologies of Information and Computer (ICFTIC) (pp. 1337-1342). IEEE.
Sunil, B. D., Sinha, I., Maheshwari, P., Todmal, S., Mallik, S., & Mishra, S. (2026). Memory Poisoning Attack and Defense on Memory Based LLM-Agents. arXiv preprint arXiv:2601.05504.
Gu, X., Yang, J., Tian, X., & Liu, M. (2025). Research on the Construction of a Human-Machine Collaborative Anti-Money Laundering System and Its Efficiency and Accuracy Enhancement in Suspicious Transaction Identification.Qiu, Y. (2024). Estimation of tail risk measures in finance: Approaches to extreme value mixture modeling. arXiv preprint arXiv:2407.05933.
Hammond, L., Chan, A., Clifton, J., Hoelscher-Obermaier, J., Khan, A., McLean, E., ... & Rahwan, I. (2025). Multi-agent risks from advanced ai. arXiv preprint arXiv:2502.14143.
Yang, Y., Leuze, C., Hargreaves, B., Daniel, B., & Baik, F. (2025). EasyREG: Easy Depth-Based Markerless Registration and Tracking using Augmented Reality Device for Surgical Guidance. arXiv preprint arXiv:2504.09498.
Tran, K. T., Dao, D., Nguyen, M. D., Pham, Q. V., O'Sullivan, B., & Nguyen, H. D. (2025). Multi-agent collaboration mechanisms: A survey of llms. arXiv preprint arXiv:2501.06322.
Alamsabi, M., Tchuindjang, M., & Brohi, S. (2026). Embedding-Based Detection of Indirect Prompt Injection Attacks in Large Language Models Using Semantic Context Analysis. Algorithms, 19(1), 92.
Bai, W., Wu, Q., Wu, K., & Lu, K. (2024). Exploring the Influence of Prompts in LLMs for Security-Related Tasks. In Workshop on Artificial Intelligence System with Confidential Computing (AISCC 2024)(San Diego, CA). USA. https://dx. doi. org/10.14722/aiscc.
Kaesberg, L. B., Becker, J., Wahle, J. P., Ruas, T., & Gipp, B. (2025, July). Voting or consensus? decision-making in multi-agent debate. In Findings of the Association for Computational Linguistics: ACL 2025 (pp. 11640-11671).
Du, Y. (2025). Research on Deep Learning Models for Forecasting Cross-Border Trade Demand Driven by Multi-Source Time-Series Data. Journal of Science, Innovation & Social Impact, 1(2), 63-70.
Brady, T. F., Robinson, M. M., Williams, J. R., & Wixted, J. T. (2023). Measuring memory is harder than you think: How to avoid problematic measurement practices in memory research. Psychonomic Bulletin & Review, 30(2), 421-449.
Liu, S., Feng, H., & Liu, X. (2025). A Study on the Mechanism of Generative Design Tools' Impact on Visual Language Reconstruction: An Interactive Analysis of Semantic Mapping and User Cognition. Authorea Preprints.
Sasikumar, A., Ravi, L., Kotecha, K., Abraham, A., Devarajan, M., & Vairavasundaram, S. (2023). A secure big data storage framework based on blockchain consensus mechanism with flexible finality. IEEE Access, 11, 56712-56725.
This work is licensed under a Creative Commons Attribution 4.0 International License.
Copyright (c) 2026 Toby Walsh, Anton van den Hengel, Stephen J. Roberts (Author)