https://arxiv.org/pdf/2504.01990#page=64.19
Memory
1 Representation
| 名称 | 年 | 引用 | ||
|---|---|---|---|---|
| Sensory | Text-based | RecAgent | 2025 | 95 |
| Sensory | Text-based | CoPS | 2024 | 29 |
| Sensory | Text-based | MemoryBank | 2024 | 300 |
| Sensory | Text-based | Memory Sandbox | 2023 | 46 |
| Sensory | Multi-modal | VideoAgent | 2024 | 88 |
| Sensory | Multi-modal | WorldGPT | 2024 | 48 |
| Sensory | Multi-modal | AgentS | 2024 | 51 |
| Sensory | Multi-modal | OS-Copilot | 2024 | 117 |
| Sensory | Multi-modal | MuLan | 2024 | 3 |
| Short-term | Context | MemGPT | 2023 | 203 |
| Short-term | Context | KARMA | 2024 | 10 |
| Short-term | Context | LSFS | 2024 | 1 |
| Short-term | Context | OSCAR | 2024 | 15 |
| Short-term | Context | RCI | 2023 | 445 |
| Short-term | Working | Generative Agent | 2023 | 2705 |
| Short-term | Working | RLP | 2023 | 17 |
| Short-term | Working | CALYPSO | 2023 | 62 |
| Short-term | Working | HiAgent | 2024 | 24 |
| Long-term | Semantic | AriGraph | 2024 | 27 |
| Long-term | Semantic | RecAgent | 2025 | 95 |
| Long-term | Semantic | HippoRAG | 2024 | 124 |
| Long-term | Episodic | MobileGPT | 2023 | 30 |
| Long-term | Episodic | MemoryBank | 2024 | 300 |
| Long-term | Episodic | Episodic Verbalization | 2024 | 6 |
| Long-term | Episodic | MrSteve | 2024 | 5 |
| Long-term | Procedural | AAG | 2024 | 1 |
| Long-term | Procedural | Cradle | 2024 | 55 |
| Long-term | Procedural | ARVIS-1 | 2024 | 121 |
| Long-term | Procedural | LARP | 2023 | 20 |
2 Lifecycle
| 名称 | 年 | 引用 | ||
|---|---|---|---|---|
| Acquisition | Information Compression | HiAgent | 2024 | 24 |
| Acquisition | Information Compression | LMAgent | 2024 | 5 |
| Acquisition | Information Compression | ReadAgent | 2024 | 39 |
| Acquisition | Information Compression | M2WF | 2025 | 2 |
| Acquisition | ExperienceConsolidation | ExpeL | 2024 | 300 |
| Acquisition | ExperienceConsolidation | MindOS | 2024/5 | 4/40 |
| Encoding | Selective Attention | AgentCorrd | 2024 | 30 |
| Encoding | Selective Attention | MS | 2024 | 19 |
| Encoding | Selective Attention | GraphVideoAgent | 2025 | 1 |
| Encoding | Selective Attention | A-MEM | 2024/5 | 6/45 |
| Encoding | Multi-modalFusion | Optimus-1 | 2024 | 41 |
| Encoding | Multi-modalFusion | Optimus-2 | 2025 | 8 |
| Encoding | Multi-modalFusion | JARVIS-1 | 2024 | 121 |
| Derivation | Reflection | Agent S | 2024 | 51 |
| Derivation | Reflection | OSCAR | 2024 | 15 |
| Derivation | Reflection | R2D2 | 2025 | 0 |
| Derivation | Reflection | Mobile-Agent-E | 2025 | 39 |
| Derivation | Summarization | SummEdits | 2023 | 72 |
| Derivation | Summarization | SCM | 2023 | 22 |
| Derivation | Summarization | Healthcare Copilot | 2024/5 | 30/59 |
| Derivation | Knowledge Distillation | Knowagent | 2024 | 56 |
| Derivation | Knowledge Distillation | AoTD | 2024 | 7 |
| Derivation | Knowledge Distillation | LDPD | 2025 | 8 |
| Derivation | Knowledge Distillation | Sub-goal Distillation | 2024 | 3 |
| Derivation | Knowledge Distillation | MAGDi | 2024 | 22 |
| Derivation | Selective Forgetting | Lyfe Agent | 2023 | 41 |
| Derivation | Selective Forgetting | TiM | 2023 | 57 |
| Derivation | Selective Forgetting | MemoryBank | 2024 | 301 |
| Derivation | Selective Forgetting | S3 | 2023/4 | 100/40 |
| Retrieval | Indexing | HippoRAG | 2024 | 126 |
| Retrieval | Indexing | TradingGPT | 2023 | 64 |
| Retrieval | Indexing | LongMemEval | 2024 | 33 |
| Retrieval | Indexing | SeCom | 2025 | 5 |
| Retrieval | Matching | Product Keys | 2019 | 161 |
| Retrieval | Matching | OSAgent | 2024 | 5/40 |
| Neural Memory | Associative Memory | Hopfield Networks | 2017/20 | 277/749 |
| Neural Memory | Associative Memory | Neural Turing Machines | 2022 | 17 |
| Neural Memory | ParameterIntegration | MemoryLLM | 2024 | 34 |
| Neural Memory | ParameterIntegration | SELF-PARAM | 2024 | 2 |
| Neural Memory | ParameterIntegration | MemoRAG | 2024 | 11 |
| Neural Memory | ParameterIntegration | TTT-Layer | 2024 | 128 |
| Neural Memory | ParameterIntegration | Titans | 2024 | 71 |
| Neural Memory | ParameterIntegration | R3Mem | 2025 | 3 |
| Utilization | RAG | RAGLAB | 2024 | 17 |
| Utilization | RAG | Adaptive Retrieval | 2022 | 681 |
| Utilization | RAG | Atlas | 2023/4 | 4/5 |
| Utilization | Long-context Modeling | RMT | 2022/3 | 208/105 |
| Utilization | Long-context Modeling | AutoCompresso | 2023 | 211 |
| Utilization | Long-context Modeling | ICAE | 2023 | 169 |
| Utilization | Long-context Modeling | Gist | 2023 | 239 |
| Utilization | Long-context Modeling | CompAct | 2024 | 27 |
| Utilization | Alleviating Hallucination | Lamini | 2024 | 11 |
| Utilization | Alleviating Hallucination | Memoria | 2023 | 7 |
| Utilization | Alleviating Hallucination | PEER | 2024 | 48/65 |
例如，RecAgent[259]采用基于llm的感觉记忆模块对原始观测进行编码，同时过滤噪声和不相关的内容。
例如，RecAgent[259]采用了一种带有重要性评分系统的注意力机制，该系统为压缩的观察值分配相关性分数，优先考虑关键输入，如特定项目的交互，同时强调不太重要的动作。
例如，RecAgent[259]通过将每个观测值与用户行为模拟环境中模拟回合的开始相对应的时间戳相关联来建模保留，该时间戳表示为⟨observation，重要性评分，时间戳⟩
在像MemoryBank[261]这样的人工智能伙伴系统中，语义记忆以自然语言构建用户画像，而情景记忆保留交互历史，增强个性化和上下文感知行为。
在更细粒度的遗忘机制中，MemoryBank[261]采用艾宾浩斯遗忘曲线（Ebbinghaus forgetting Curve）来量化遗忘率，同时考虑了时间衰减和间隔效应，即重新学习信息比第一次学习更容易的原则。
Expel[96]构建了一个经验库，从训练任务中收集和提取见解，促进对未见任务的推广。
ExpeL[96]利用反思来收集过去的经验，以便将其推广到看不见的任务，并支持失败后的反复尝试。
通过像reflex[75]和ExpeL[96]这样的系统，智能体通过自主管理经验收集、分析和应用的完整周期，实现了复杂的体验式学习，使它们能够从成功和失败中有效地学习。