Lifelong Learning of Large Language Model Based Agents: A Roadmap
The DOI matches its Crossref record; 218 of the 218 references it cites were found in a registry or an index.
Abstract
Lifelong learning, also known as continual or incremental learning, is a crucial component for advancing Artificial General Intelligence (AGI) by enabling systems to continuously adapt in dynamic environments. While large language models (LLMs) have demonstrated impressive capabilities in natural language processing, existing LLM agents are typically designed for static systems and lack the ability to adapt over time in response to new challenges. This survey is the first to systematically summarize the potential techniques for incorporating lifelong learning into LLM-based agents. We categorize the core components of these agents into three modules: the perception module for multimodal input integration, the memory module for storing and retrieving evolving knowledge, and the action module for grounded interactions with the dynamic environment. We highlight how these pillars collectively enable continuous adaptation, mitigate catastrophic forgetting, and improve long-term performance. This survey provides a roadmap for researchers and practitioners working to develop lifelong learning capabilities in LLM agents, offering insights into emerging trends, evaluation metrics, and application scenarios.
Overview
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
In the citing paper
“Empirically, this kind of reviewer setup has been shown to reduce hallucinations in LLM-generated answers [47,48].”
From A Review of Multi-Agent AI Systems for Biological and Clinical Data Analysis (2026) · cites this paper as [47]
In the citing paper
“Zheng et al. (2025) outlined strategies for incorporating lifelong learning into agent architecture by dividing the system into a perception module (to handle new modalities), a memory module (to accumulate evolving knowledge), and an action module (to adapt behaviors) [47].”
From A Review of Multi-Agent AI Systems for Biological and Clinical Data Analysis (2026) · cites this paper as [47]
In the citing paper
“An agent adopting a poorly designed memory architecture can suffer from factual contradictions, catastrophic forgetting, or unacceptable latencies during continuous execution [6, 38].”
From Are We Ready For An Agent-Native Memory System? (2026) · cites this paper as [38]
Citations
Cites
Items 201–218 of 218
JARVIS-1: Open-World Multi-task Agents with Memory-Augmented Multimodal Language Models
Wang, Z., Cai, S., Liu, A., Jin, Y., Hou, J., Zhang, B., Lin, H., He, Z., Zheng, Z., Yang, Y., et al. (2023) · arXiv · doi:10.48550/arxiv.2311.05997 · Cited by 6
MemoChat: Tuning LLMs to Use Memos for Consistent Long-Range Open-Domain Conversation
Lu, J., An, S., Lin, M., Pergola, G., He, Y., Yin, D., Sun, X., & Wu, Y. (2023) · arXiv · doi:10.48550/arxiv.2308.08239 · Cited by 6
EPO: Hierarchical LLM Agents with Environment Preference Optimization
Zhao, Q., Fu, H., Sun, C., & Konidaris, G. (2024) · arXiv · doi:10.48550/arxiv.2408.16090 · Cited by 6
Wang, Z. Z., Mao, J., Fried, D., & Neubig, G. (2024) · arXiv · doi:10.48550/arxiv.2409.07429 · Cited by 5
H-MEM: Hierarchical Memory for High-Efficiency Long-Term Reasoning in LLM Agents
Sun, H., Zeng, S., & Zhang, B. (2026) · Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers) · doi:10.18653/v1/2026.eacl-long.15 · Cited by 4
EASYTOOL: Enhancing LLM-based Agents with Concise Tool Instruction
Yuan, S., Song, K., Chen, J., Tan, X., Shen, Y., Kan, R., Li, D., & Yang, D. (2024) · arXiv · doi:10.48550/arxiv.2401.06201 · Cited by 4
Math Agents: Computational Infrastructure, Mathematical Embedding, and Genomics
Swan, M., Kido, T., Roland, E., & Santos, R. P. D. (2023) · arXiv · doi:10.48550/arxiv.2307.02502 · Cited by 4 · cited 1×
“In education [171], [172] and healthcare [173], agents adaptively integrate new knowledge to provide personalized teaching and decision support in safety-critical contexts.”
Navigating the Digital World as Humans Do: Universal Visual Grounding for GUI Agents
Gou, B., Wang, R., Zheng, B., Xie, Y., Chang, C., Shu, Y., Sun, H., & Su, Y. (2024) · arXiv · doi:10.48550/arxiv.2410.05243 · Cited by 3
Enhance Lifelong Model Editing with Continuous Data-Adapter Association
Li, J., Wang, Q., Wang, Z., Zhang, Y.-D., & Mao, Z.-D. (2024) · arXiv.org · doi:10.48550/arxiv.2408.11869 · Cited by 2
LongAgent: Scaling Language Models to 128k Context through Multi-Agent Collaboration
Zhao, J., Zu, C., Xu, H., Lu, Y., He, W., Ding, Y., Gui, T., Zhang, Q., & Huang, X. (2024) · arXiv · doi:10.48550/arxiv.2402.11550 · Cited by 2
Tree Search for Language Model Agents
Koh, J. Y., McAleer, S., Fried, D., & Salakhutdinov, R. (2024) · arXiv · doi:10.48550/arxiv.2407.01476 · Cited by 2
Xu, J., Guo, Z., Hu, H., Chu, Y., Wang, X., He, J., Wang, Y., Shi, X., He, T., Zhu, X., et al. (2025) · arXiv · doi:10.48550/arxiv.2509.17765 · Cited by 2
Position: Episodic Memory is the Missing Piece for Long-Term LLM Agents
Pink, M., Wu, Q., Vo, V. A., Turek, J., Mu, J., Huth, A., & Toneva, M. (2025) · arXiv · doi:10.48550/arxiv.2502.06975 · Cited by 1
SWE-Bench-CL: Continual Learning for Coding Agents
Joshi, T., Chowdhury, S., & Uysal, F. (2025) · arXiv · Cited by 1
SteP: Stacked LLM Policies for Web Actions
Sodhi, P., Branavan, S. R. K., Artzi, Y., & McDonald, R. (2023) · arXiv · doi:10.48550/arxiv.2310.03720 · Cited by 1
LifelongAgentBench: Evaluating LLM Agents as Lifelong Learners
Zheng, J., Cai, X., Li, Q., Zhang, D., Li, Z., Zhang, Y., Song, L., & Ma, Q. (2025) · arXiv · doi:10.48550/arxiv.2505.11942
Self-citationN-Critics: Self-Refinement of Large Language Models with Ensemble of Critics
Mousavi, S., Gutiérrez, R. L., Rengarajan, D., Gundecha, V., Babu, A. R., Naug, A., Guillen, A., & Sarkar, S. (2023) · arXiv · doi:10.48550/arxiv.2310.18679
COLT: Enhancing Video Large Language Models with Continual Tool Usage
Liu, Y. (2026) · MBZUAI iRep · doi:10.82419/399
Research works printed but not linked: 5
- Zhou, 2024, WebArena: A realistic web environment for building autonomous agents
- Ding, 2020, CogLTX: Applying bert to long texts, 33, 12792
- Lu, 2021, Container: Context aggregation network, 34, 1960
- Zhang, 2024, CPPO: Continual learning for reinforcement learning with human feedback
- Raman, 2022, Planning with large language models via corrective re-prompting
Cited by
A Survey of Agent Memory in the Second Half: Towards Self-Evolving and Long-Horizon Agents
Huang, W.-C., Zhang, W., Liang, Y., Bei, Y., Chen, Y., Feng, T., Pan, X., Tan, Z., Wang, Y., Wei, T., et al. (2026) · arXiv · doi:10.48550/arxiv.2602.06052 · Cited by 70
From System 1 to System 2: A Survey of Reasoning Large Language Models
Zhang, D., Li, Z., Zhang, M.-L., Zhang, J., Liu, Z., Yao, Y., Xu, H., Zheng, J., Chen, X., Zhang, Y., et al. (2025) · IEEE Transactions on Pattern Analysis and Machine Intelligence · doi:10.1109/tpami.2025.3637037 · Cited by 612
“However, standard LLMs are prone to catastrophic forgetting, where updates for new tasks overwrite previous capabilities, and often struggle with domain shifts in reasoning contexts [244].”
Self-citationDeep Reinforcement Learning in the Era of Foundation Models: A Survey
Mienye, I. D., Esenogho, E., & Modisane, C. (2026) · Computers · doi:10.3390/computers15010040 · Cited by 13
A Holistic Review of Agentic AI Frameworks, Applications, and Research Trajectories
Peykani, P., Ghanidel, S., Javadi-Sisi, I., Snåšel, V., & Mirjalili, S. (2026) · Archives of Computational Methods in Engineering · doi:10.1007/s11831-026-10675-8 · Cited by 7
A Review of Multi-Agent AI Systems for Biological and Clinical Data Analysis
Spieser, J., Balapour, A., Meller, J., Patra, K. C., & Shamsaei, B. (2026) · Methods and Protocols · doi:10.3390/mps9020033 · Cited by 5
“Empirically, this kind of reviewer setup has been shown to reduce hallucinations in LLM-generated answers [47,48].”
Influential
Metadata
- Authors
- Zheng, J., Shi, C., Cai, X., Li, Q., Zhang, D., Li, C., Yu, D., & Ma, Q.
- Affiliations
- South China University of Technology (CN)
- Mohamed bin Zayed University of Artificial Intelligence (AE)
- Tencent (China) (CN)
- Tencent AI Lab, Bellevue, WA, USA
- Venue
- IEEE Transactions on Pattern Analysis and Machine Intelligence
- Published
- 2026-05
- Volume
- 48
- Issue
- 5
- Pages
- 5552-5571
- ISSN
- 0162-8828 (ISSN-L)
- 2160-9292
- 1939-3539
- OpenAlex
- W7118167202
- Cited by
- 7
- References
- 98
- Citations by year
- 2026: 5
- FWCI
- 4.77
- Type
- journal article
- Language
- en
- Open access
- closed
- Licence
- Locations
- IEEE Transactions on Pattern Analysis and Machine Intelligence (publishedVersion)
- PubMed (publishedVersion)
- Funders
- National Natural Science Foundation of China (62272173, 62273109)
- Natural Science Foundation of Guangdong Province (2024A1515010089, 2022A1515010179)
- Science and Technology Planning Project of Guangdong Province (2023A0505050106)
- National Key R&D Program of China (2023YFA1011601)
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Domain Adaptation and Few-Shot Learning
- Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
- Physical Sciences › Computer Science › Artificial Intelligence › AI-based Problem Solving and Planning
- Keywords
- lifelong learning, continual learning, LLM agents, artificial general intelligence, catastrophic forgetting, memory modules, multisensory perception, knowledge retrieval, long-term performance
- MeSH terms
- Large Language Models, Artificial Intelligence, Humans, Language, Learning, Natural Language Processing
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.