A Study on the Coupling Relationship Between AI Personalization and Long-Context Management
Main Article Content
Keywords
large language model, personalization, long context, agent memory, user preference, coupling relationship
Abstract
With the large-scale deployment of large language models (LLMs) in various intelligent assistants and agent systems, homogenized outputs and the loss of user preferences across sessions have become key bottlenecks restricting improvements in user experience. Personalization technology aims to adapt LLMs to the language habits, thinking patterns, and subjective preferences of different users, while long-context and agent memory management addresses the underlying technical problems of storing, filtering, and reusing temporal information for LLMs. Current academic research mostly treats the two directions as independent fields, and systematic reviews of their internal coupling relationship are relatively scarce. This study first defines the core concepts of LLM personalization and long-context memory management and systematically reviews the technical principles, advantages, and existing limitations of four personalization paradigms: prompt-engineering personalization, lightweight parameter fine-tuning, user-preference alignment, and dynamic user profiling. It then divides long-context and agent-memory technologies into five routes: native-window expansion, context pruning and compression, adaptive context extraction, hierarchical memory architectures, and external retrieval-augmented memory, and analyzes the applicable scenarios and technical shortcomings of each route. On this basis, this study reviews progress in four intersecting directions: preference-aware memory updating, structured persona-memory representation, long-horizon personalization evaluation benchmarks, and training-free collaborative control at inference time. It further analyzes the internal coupling mechanism through which the two directions empower each other and summarizes three core conflicts: the capacity-accuracy trade-off, the conflict between general governance rules and personalized characteristics, and the conflict between computational cost and personalization stability. The study also summarizes five research challenges: difficulty in adapting to the dynamic evolution of user preferences, failure to recall implicit preferences, memory errors and preference drift, fragmented evaluation systems, and difficulty balancing engineering implementation costs. It proposes subsequent research paths in six dimensions: cognitively inspired memory architectures, preference-aware context compression algorithms, multimode hybrid indexing and retrieval, unified long-horizon personalization evaluation protocols, memory-error verification mechanisms, and multimodal long-term personalization. The findings indicate that long-term personalized intelligent systems cannot optimize the personalization module or memory module in isolation. Their coupling relationship should serve as the underlying design logic to promote the iterative upgrading of LLMs toward adaptation to individual user characteristics and to provide a reference for the theoretical research and engineering development of long-term agents and personalized dialogue assistants.
References
- [1]Zhao M, Xu Y, Chen Q, et al. Toward Personalized LLM-Powered Agents: Foundations, Evaluation, and Future Directions[J]. arXiv preprint arXiv:2602.22680, 2026.
- [2]Liu X, Chen Y. Personalization of Large Language Models: A Survey[J]. arXiv preprint arXiv:2411.0027, 2024
- [3]Li Y, Wang Z. A Survey of Personalized Large Language Models: Progress and Future Directions[J]. arXiv preprint arXiv:2502.11528, 2025
- [4]Luo J, Zhang Y. Large Language Model Agent: A Survey on Methodology, Applications and Challenges[J]. arXiv preprint arXiv:2503.21460, 2025
- [5]Xu, P., Singh, A., & Choi, Y. (2025). LaMP-2: Benchmarking Long-Term Personalization for Large Language Models [Unverified preprint metadata; source manuscript lists arXiv:2501.08741].
- [6]Kumar I, Viswanathan S, Yerra S, et al. Know It, Act on It: Investigating Memory Utilization in LLM Personalization[J]. arXiv preprint arXiv:2607.29433, 2026
- [7]Nie, C., Fu, C., Zhang, Y., et al. (2026). PersonaVLM: Long-Term Personalized Multimodal LLMs [Unverified preprint metadata; source manuscript lists arXiv:2604.13074].
- [8]Chhikara, P., Khant, D., Aryan, S., Singh, T., & Yadav, D. (2025). Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2504.19413
- [9]Huang, J., & Li, Z. (2024). A Survey on Long-Context Large Language Models [Unverified preprint metadata; source manuscript lists arXiv:2406.09128].
- [10]Gao, F., & Yang, Z. (2025). Long-Context, Less Focus: Understanding Information Dilution in LLMs [Unverified preprint metadata; source manuscript lists arXiv:2502.13572].
- [11]Maharana, A., Lee, D. H., Tulyakov, S., Bansal, M., Barbieri, F., & Fang, Y. (2024). Evaluating Very Long-Term Conversational Memory of LLM Agents [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2402.17753
- [12]Sun, T., & Zhang, H. (2024). Context Pruning Methods for Large Language Models: A Survey [Unverified preprint metadata; source manuscript lists arXiv:2409.08391].
- [13]Zhang, S., & Liu, H. (2024). Dynamic User Profiling for Continual LLM Personalization [Unverified preprint metadata; source manuscript lists arXiv:2408.15217].
- [14]Chen, L., & Zhou, H. (2024). Lightweight Personalization for LLMs: Prompt, LoRA and RAG Comparison [Unverified preprint metadata; source manuscript lists arXiv:2410.02744].
- [15]Wu, X., & Zhou, R. (2025). Dynamic User Profile Extraction from Multi-Session Dialogues [Unverified preprint metadata; source manuscript lists arXiv:2502.09597].
- [16]Peng, B., & Lin, Y. (2025). A Review of Long-Context Techniques for Large-Language-Model-Based Agents [Unverified preprint metadata; source manuscript lists arXiv:2504.02381].
- [17]Han, Z., & Wang, J. (2025). Memory Mechanisms for LLM Agents: Techniques and Evaluations [Unverified preprint metadata; source manuscript lists arXiv:2507.02203].
- [18]Wu, Z., & Zhang, X. (2024). Survey on Retrieval-Enhanced Memory for Conversational Agents [Unverified preprint metadata; source manuscript lists arXiv:2412.04254].
- [19]Ding, S., & Qian, Y. (2025). Risk of Erasing User Preference by Generic Context Pruning [Unverified preprint metadata; source manuscript lists arXiv:2512.02748].
- [20]Liu H, Wang J. Preference-Aware Memory Update for Long-Term LLM Agents[J]. arXiv preprint arXiv:2510.09720, 2025
- [21]Zhang Y, Wang H. From Storage to Experience: A Survey on the Evolution of LLM Agent Memory Mechanisms[J]. arXiv preprint arXiv:2605.06716, 2026
- [22]Jiang, T., & Li, S. (2024). Challenges in Real-World LLM Personalization [Unverified preprint metadata; source manuscript lists arXiv:2405.13219].
- [23]Madaan, A., Tandon, N., Gupta, P., et al. (2022). MemPrompt: Memory-Assisted Prompting for Large Language Models [Unverified preprint metadata; source manuscript lists arXiv:2210.11797].
- [24]Salemi, A., Kim, S., & Schölkopf, B. (2023). Prompt Compression for Large Language Models [Unverified preprint metadata; source manuscript lists arXiv:2305.14821].
- [25]Shi, W., & Liu, J. (2025). Token Budget Management for Personalized System Prompts [Unverified preprint metadata; source manuscript lists arXiv:2503.12774].
- [26]Lin, Y., & Zhang, Q. (2024). User-Profile Summarization for LLM Personalization [Unverified preprint metadata; source manuscript lists arXiv:2412.08152].
- [27]Zhou, K., & Wu, L. (2025). Abstractive Compression of User Interaction History for Personalized Agents [Unverified preprint metadata; source manuscript lists arXiv:2507.11364].
- [28]Hu, E. J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., & Chen, W. (2021). LoRA: Low-Rank Adaptation of Large Language Models [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2106.09685
- [29]Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., de Laroussilhe, Q., Gesmundo, A., Attariyan, M., & Gelly, S. (2019). Parameter-Efficient Transfer Learning for NLP [Preprint]. arXiv. https://doi.org/10.48550/arXiv.1902.00751
- [30]Li, X. L., & Liang, P. (2021). Prefix-Tuning: Optimizing Continuous Prompts for Generation [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2101.00190
- [31]Zhang, R., & Liu, Y. (2025). Overfitting Analysis for User-Specific LoRA Personalization [Unverified preprint metadata; source manuscript lists arXiv:2505.02873].
- [32]Wang, Z., & Chen, M. (2024). Data Scarcity in Individual User LLM Personalization [Unverified preprint metadata; source manuscript lists arXiv:2407.16291].
- [33]Lee, J., & Park, S. (2025). Incremental LoRA for Continual User Preference Update [Unverified preprint metadata; source manuscript lists arXiv:2508.15623].
- [34]Zhao, H., & Li, W. (2024). Catastrophic Forgetting in Continual LLM Personalization [Unverified preprint metadata; source manuscript lists arXiv:2411.08476].
- [35]Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C. L., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., Schulman, J., Hilton, J., Kelton, F., Miller, L., Simens, M., Askell, A., Welinder, P., Christiano, P., Leike, J., & Lowe, R. (2022). Training language models to follow instructions with human feedback [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2203.02155
- [36]Rafailov, R., Sharma, A., Mitchell, E., Ermon, S., Manning, C. D., & Finn, C. (2023). Direct Preference Optimization: Your Language Model is Secretly a Reward Model [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2305.18290
- [37]Xie, Z., Wu, J., Shen, Y., Xia, Y., Li, X., Chang, A., Rossi, R., Kumar, S., Majumder, B. P., Shang, J., Ammanabrolu, P., & McAuley, J. (2025). A Survey on Personalized and Pluralistic Preference Alignment in Large Language Models [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2504.07070
- [38]Liu, Y., & Yang, B. (2024). User-Specific DPO for Personalized Large Language Models [Unverified preprint metadata; source manuscript lists arXiv:2406.17819].
- [39]Zhang, H., & Wang, Y. (2025). Meta-Alignment for Fast User Adaptation of LLMs [Unverified preprint metadata; source manuscript lists arXiv:2501.04211].
- [40]Tan, S., & Li, Q. (2025). Decay of Preference Alignment Over Extended Dialogue Sessions [Unverified preprint metadata; source manuscript lists arXiv:2509.14726].
- [41]Chen, J., & Gong, L. (2025). Inference-Time Personalized Alignment Without Per-User Tuning [Unverified preprint metadata; source manuscript lists arXiv:2506.03948].
- [42]Wang, P., & Zeng, H. (2024). Preference-Guided Decoding for Zero-Shot LLM Personalization [Unverified preprint metadata; source manuscript lists arXiv:2410.14635].
- [43]Wang, T., Tao, M., Fang, R., Wang, H., Wang, S., Jiang, Y. E., & Zhou, W. (2024). AI PERSONA: Towards Life-long Personalization of LLMs [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2412.13103
- [44]Zheng, Y., & Guo, S. (2026). Evolving User Persona for Long-Term LLM Agents [Unverified preprint metadata; source manuscript lists arXiv:2601.12587].
- [45]Yu, L., & Zhang, T. (2025). Detection of Out-of-Date Preferences in Dynamic Profiles [Unverified preprint metadata; source manuscript lists arXiv:2511.05472].
- [46]Meng, X., & He, K. (2024). Profile Distortion Caused by Context Window Truncation in User Modeling [Unverified preprint metadata; source manuscript lists arXiv:2408.19762].
- [47]Feng, Q., & Sun, B. (2026). Benchmark for Dynamic User Profile Evolution [Unverified preprint metadata; source manuscript lists arXiv:2603.11892].
- [48]Beltagy, I., Peters, M. E., & Cohan, A. (2020). Longformer: The Long-Document Transformer [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2004.05150
- [49]Zaheer, M., Guruganesh, G., Dubey, A., Ainslie, J., Alberti, C., Ontanon, S., Pham, P., Ravula, A., Wang, Q., Yang, L., & Ahmed, A. (2020). Big Bird: Transformers for Longer Sequences [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2007.14062
- [50]Ding, J., & Zhou, W. (2025). Sparse Attention Variants for Extending LLM Context Windows [Unverified preprint metadata; source manuscript lists arXiv:2503.07415].
- [51]Chen, Y., & Zhang, L. (2025). Empirical Study of Attention Dilution in Million-Token LLM Contexts [Unverified preprint metadata; source manuscript lists arXiv:2505.08917].
- [52]Sui, X., & Liu, Z. (2024). When Bigger Context Window Does Not Mean Better Performance for LLMs [Unverified preprint metadata; source manuscript lists arXiv:2410.17123].
- [53]Kim, N., & Cho, J. (2023). Context Pruning by Importance Scoring for Long-Document LLMs [Unverified preprint metadata; source manuscript lists arXiv:2307.06904].
- [54]Xu, J., & Yu, H. (2024). Sliding-Window Abstractive Summarization for Long Dialogue History [Unverified preprint metadata; source manuscript lists arXiv:2404.05911].
- [55]Yang, M., & Qian, R. (2025). Hybrid Vector-Text Compression for Agent Conversation Context [Unverified preprint metadata; source manuscript lists arXiv:2506.18422].
- [56]Wang X, Zhou L. VitaBench 2.0: Evaluating Personalized and Proactive Agents in Long-Term User Interactions[J]. arXiv preprint arXiv:2605.27141, 2026
- [57]Ji, S., Li, Y., & Hooi, B. (2026). Memory is Reconstructed, Not Retrieved: Graph Memory for LLM Agents [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2606.06036
- [58]Zhang, N., & Yang, X. (2026). HiMem: Hierarchical Long-Term Memory for LLM Long-Horizon Agents [Unverified preprint metadata; source manuscript lists arXiv:2602.10147].
- [59]Zhong, W., Guo, L., Gao, Q., Ye, H., & Wang, Y. (2023). MemoryBank: Enhancing Large Language Models with Long-Term Memory [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2305.10250
- [60]Wang R, Lu H. AdaMem: Adaptive User-Centric Memory for Long-Horizon Dialogue Agents[J]. arXiv preprint arXiv:2603.16496, 2026
Article Sidebar
This work is licensed under a Creative Commons Attribution 4.0 International License.