Continual Learning 记忆管理（Memory Management）相关文献整理

按方法类型分类，并标注发表会议/期刊及年份 · 整理时间 2026-07-01

📚 综述类文献

系统梳理持续学习中记忆/回放机制的综述与分类框架

年份	论文	会议/期刊
2024	Class-Incremental Learning: A Survey 涵盖 exemplar-based memory 方法分类	TPAMI
2021	A Continual Learning Survey: Defying Forgetting in Classification Tasks	TPAMI
2021	Online Continual Learning in Image Classification: An Empirical Survey	Neurocomputing
2020	Continual Lifelong Learning in Natural Language Processing: A Survey 提出 Rehearsal / Pseudo-rehearsal / Regularization / Memory 四分类	COLING
2020	Class-Incremental Learning: Survey and Performance Evaluation	arXiv
2019	Continual Lifelong Learning with Neural Networks: A Review	Neural Networks
2019	Three Scenarios for Continual Learning	arXiv
2026	Survey on AI Memory 面向 LLM Agent 的 4W 记忆分类体系（记什么/何时记/如何检索/如何评估）	arXiv (BAI-LAB)

🗄️ 经验回放 / 样本缓冲区管理（Rehearsal, Exemplar & Buffer Selection）Replay

存储/精选历史真实样本，核心问题是「存哪些样本」「如何检索用于回放」

年份	论文	会议
2017	iCaRL: Incremental Classifier and Representation Learning 羊群效应(herding)优先样本选择	CVPR
2017	Gradient Episodic Memory for Continual Learning (GEM)	NeurIPS (NIPS)
2019	Efficient Lifelong Learning with A-GEM	ICLR
2019	Gradient based Sample Selection for Online Continual Learning (GSS)	NeurIPS
2019	Experience Replay for Continual Learning (CLEAR)	NeurIPS
2019	Online Continual Learning with Maximally Interfered Retrieval (MIR) 按干扰程度检索回放样本	NeurIPS
2021	Rainbow Memory: Continual Learning with a Memory of Diverse Samples	CVPR
2021	RMM: Reinforced Memory Management for Class-Incremental Learning 强化学习动态分配 exemplar 内存预算	NeurIPS
2021	Gradient-based Editing of Memory Examples for Online Task-free Continual Learning	NeurIPS
2021	Information-theoretic Online Memory Selection for Continual Learning	ICLR
2021	Rehearsal Revealed: The Limits and Merits of Revisiting Samples in Continual Learning	ICCV
2021	Continual Learning on Noisy Data Streams via Self-Purified Replay	ICCV
2022	Online Coreset Selection for Rehearsal-based Continual Learning	ICLR
2022	Memory Replay with Data Compression for Continual Learning 压缩存储降低缓冲区开销	ICLR
2023	GCR: Gradient Coreset based Replay Buffer Selection for Continual Learning	CVPR

🎨 生成式回放 / 伪回放（Generative & Pseudo-Rehearsal）Generative

不存储真实样本，用生成模型（GAN/VAE/Diffusion）合成旧任务数据作为「记忆」

年份	论文	会议
1995	Catastrophic Forgetting, Rehearsal and Pseudorehearsal 伪回放思想起源	Connection Science
2017	Continual Learning with Deep Generative Replay 开创 Deep Generative Replay 框架	NeurIPS
2018	Memory Replay GANs: Learning to Generate New Categories without Forgetting	NeurIPS
2018	Generative Replay with Feedback Connections as a General Strategy for Continual Learning	arXiv
2020	Brain-Inspired Replay for Continual Learning with Artificial Neural Networks	Nature Communications
2020	GAN Memory with No Forgetting	NeurIPS
2021	Always Be Dreaming: A New Approach for Data-Free Class-Incremental Learning	ICCV
2022	Learning to Imagine: Diversify Memory for Incremental Learning Using Unlabeled Data	CVPR
2023	DDGR: Continual Learning with Deep Diffusion-based Generative Replay 扩散模型作为记忆生成器	ICML
2023	Selective Amnesia: A Continual Learning Approach to Forgetting in Deep Generative Models EWC + 生成式回放实现选择性遗忘	NeurIPS
2024	SDDGR: Stable Diffusion-based Deep Generative Replay for Class Incremental Object Detection	CVPR
2024	Online Task-Free Continual Generative and Discriminative Learning via Dynamic Cluster Memory	CVPR

📐 梯度约束型记忆（Gradient-Constrained Episodic Memory）Gradient

用少量存储样本构造梯度约束/参考方向，防止新任务梯度更新破坏旧知识

年份	论文	会议
2017	Gradient Episodic Memory for Continual Learning (GEM)	NeurIPS
2019	Efficient Lifelong Learning with A-GEM 用平均梯度替代二次规划，提升效率	ICLR
2021	Flattening Sharpness for Dynamic Gradient Projection Memory Benefits Continual Learning	NeurIPS
2022	Continual Learning with Recursive Gradient Optimization	ICLR
2022	TRGP: Trust Region Gradient Projection for Continual Learning	ICLR

🧠 双记忆 / 互补学习系统（Dual Memory & Complementary Learning Systems）Dual-Memory

受大脑海马体-新皮层理论启发，构建短期/长期或快/慢双记忆模块协同工作

年份	论文	会议
2021	DualNet: Continual Learning, Fast and Slow	NeurIPS
2022	Learning Fast, Learning Slow: A General Continual Learning Method based on Complementary Learning System (CLS-ER) 短期/长期语义记忆双模型 + 情景记忆	ICLR
2021	Posterior Meta-Replay for Continual Learning	NeurIPS
2019	Complementary Learning for Overcoming Catastrophic Forgetting Using Experience Replay	IJCAI

🚫 免回放 / 架构与正则化记忆方法（Rehearsal-Free: Architecture & Regularization）Rehearsal-Free

不显式存储历史样本，通过参数正则化、任务专属子网络或 Prompt 池承载「记忆」（隐私更友好）

年份	论文	会议
2023	CODA-Prompt: COntinual Decomposed Attention-based Prompting for Rehearsal-Free Continual Learning	CVPR
2023	Generating Instance-level Prompts for Rehearsal-free Continual Learning (DAP)	ICCV
2024	Generating Prompts in Latent Space for Rehearsal-free Continual Learning	ACM MM
2024	Rehearsal-Free Modular and Compositional Continual Learning for Language Models	arXiv
2022	Model Zoo: A Growing Brain That Learns Continually 动态扩展网络结构承载知识	ICLR
2021	DER: Dynamically Expandable Representation for Class Incremental Learning	CVPR

🤖 LLM / Agent 记忆管理（新兴方向）LLM-Agent

面向大模型 Agent 的持续学习/记忆管理，从「参数级」记忆转向「外部结构化」记忆，是当前（2025-2026）研究热点，尚未形成统一的顶会体系，多以 arXiv 预印本 + 工程实践形式出现

年份	论文/项目	来源
2026	Survey on AI Memory 北邮 BAI-LAB，提出单/多智能体记忆架构与评估体系	arXiv
2026	Complementary RL: 让 AI 智能体像人脑一样积累和运用经验阿里巴巴 & 港科大，经验记忆 + 强化学习训练	arXiv (2603.17621)
2026	TencentDB Agent Memory 腾讯云开源，短期记忆压缩 + 长期个性化记忆	工程实践/开源项目
2026	Cloudflare Agent Memory 托管式跨会话记忆服务，结构化记忆抽取	工程实践

⚠️ LLM/Agent 记忆管理领域演进极快，上述条目多为工程博客/预印本而非同行评审顶会论文，建议结合具体研究需求进一步跟踪 arXiv 最新提交。