Adapting Without Forgetting — A Lifelong Learning Roadmap for LLM Agents

Adapting Without Forgetting

A Lifelong Learning Roadmap for LLM Agents — how agents that never update their weights still find ways to forget, and what a 2025 survey does and doesn't prove about fixing it.

arXiv:2501.07278 Zheng, Shi, Cai, Li, Zhang, Li, Yu, Ma · 2025 South China University of Technology · MBZUAI · Tencent AI Lab

The survey welds two lineages together — internal-knowledge LLM surveys and agent-architecture surveys — into one loop: perception, memory, and action, formalized as a goal-conditioned POMDP. Click a memory block to see how each type is actually used.

Perception → Memory → Action loop

Click a memory block in the diagram above — Working, Episodic, Semantic, or Parametric — to see how it's used in practice.

Mock retrieval-activation intensity across a 10-task episode, by memory type. Episodic memory climbs as an agent's skill library grows (the Voyager pattern); parametric memory stays cold except when a knowledge-editing pass fires. Hover a cell for the story behind the number.

Memory activation by task step

Freezing the backbone doesn't remove the stability-plasticity dilemma — it relocates it. Toggle between an EWC-style protected update and an open update to see which weights move, then compare what that tradeoff does to performance across a task sequence.

Weight update pattern — Stability Mode

Average Performance (AP) across a 10-task sequence

The survey inherits five formal metrics from classical continual learning — AP, AIP, FGT, BWT, FWT — but the systems it cites report free-form success rates instead. A taxonomy checkmark is not evidence a system was ever tested for forgetting.

of ~360 cited works have a table where the survey computes FGT or BWT for a named system. Confucius, ART, and GITM get the same taxonomy checkmark as papers actually designed to measure forgetting.

Taxonomy membership vs. metrics actually reported

References

1.Lifelong Learning of Large Language Model based Agents: A Roadmap — Junhao Zheng, Chengming Shi, Xidi Cai, Qiuke Li, Duzhen Zhang, Chenxing Li, Dong Yu, Qianli Ma, 2025 — arxiv.org/abs/2501.07278

2.Overcoming Catastrophic Forgetting in Neural Networks — James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, et al. (DeepMind), 2017 — Google Scholar

3.Continual Lifelong Learning with Neural Networks: A Review — German I. Parisi, Ronald Kemker, Jose L. Part, Christopher Kanan, Stefan Wermter, 2019 — Google Scholar

4.Generative Agents: Interactive Simulacra of Human Behavior — Joon Sung Park, Joseph O'Brien, Carrie J. Cai, Meredith Ringel Morris, Percy Liang, Michael S. Bernstein (Stanford / Google), 2023 — Google Scholar

5.Voyager: An Open-Ended Embodied Agent with Large Language Models — Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, Anima Anandkumar (NVIDIA, Caltech, UT Austin), 2023 — Google Scholar

6.Towards Lifelong Learning of Large Language Models: A Survey — J. Zheng, S. Qiu, C. Shi, Q. Ma, 2024 — Google Scholar

7.Loss of Plasticity in Deep Continual Learning — S. Dohare, J. F. Hernandez-Garcia, Q. Lan, P. Rahman, A. R. Mahmood, R. S. Sutton, 2024 — Google Scholar

8.A Survey on Large Language Model Based Autonomous Agents — L. Wang, C. Ma, X. Feng, et al., 2024 — Google Scholar

9.WISE: Rethinking the Knowledge Memory for Lifelong Model Editing of Large Language Models — P. Wang, Z. Li, N. Zhang, Z. Xu, Y. Yao, Y. Jiang, P. Xie, F. Huang, H. Chen, 2024 — Google Scholar

10.Catastrophic Interference in Connectionist Networks: The Sequential Learning Problem — M. McCloskey, N. J. Cohen, 1989 — Google Scholar