记忆综述 - 百家AI

psychology 统一理论框架：连接认知心理学与计算科学 Unified Framework: Psychology Meets Computing

综述从认知心理学与神经科学汲取灵感，将人类记忆模型映射为 AI 架构原则，帮助解释“为何这样设计”而不仅是“怎么堆模块”。

The survey draws inspiration from psychology and neuroscience to map human memory to AI architecture, explaining the "Why" behind the design.

Atkinson–Shiffrin 分层记忆Atkinson–Shiffrin Model：启发从瞬时缓存到长期存储的分层设计。Inspires hierarchical design from cache to long-term storage.
工作记忆模型Working Memory Model：支持构建多组件、可动态读写的“工作空间”。Supports multi-component, dynamic "workspaces."
互补学习系统（CLS）Complementary Learning Systems：强调“快速获取新知识”与“缓慢巩固知识”的协同，缓解记忆更新中的稳定性–可塑性矛盾。Balances fast acquisition and slow consolidation to ease the stability-plasticity dilemma.

schema 独创的 4W 记忆分类法：把“碎片研究”放进同一坐标系 Original 4W Taxonomy: Unifying Scattered Research

用四个维度统一定义内存机制，便于对比不同方案，也便于你在工程落地时做取舍。

Defining memory mechanisms through four dimensions for better comparison and engineering trade-offs.

          When（生命周期）When (Lifecycle)：Transient → Session → Persistent。

                What（内存类型）What (Memory Type)：
                程序性 / 陈述性 / 元认知 / 个性化。
                Procedural / Declarative / Metacognitive / Personalized.
            
                How（存储形式）How (Format)：
                参数化隐式（权重/隐藏态） vs 显式（文本/向量/知识图谱）。
                Implicit (Weights) vs. Explicit (Text/Vector/KG).
            
                Which（模态维度）Which (Modality)：
                文本单模态 → 视听等多模态融合。
                Text-only → Multi-modal Fusion.

groups 单智能体 vs. 多智能体：从个体记忆到群体智能 Single vs. Multi-Agent: Individual to Collective

记忆机制随系统规模演进：单体侧重突破上下文窗口与自我进化，多体侧重共享与协作。

Memory evolves with scale: Single agents focus on context windows; multi-agents focus on collaboration.

单智能体Single Agent：分层（Hierarchical）与类 OS（OS-like）设计，强化长期一致性与可扩展记忆。Hierarchical and OS-like designs for enhanced long-term consistency and scalable memory.
多智能体（MAS）Multi-Agent (MAS)：关注通信（显式符号 vs 隐式隐空间）与共享（任务级/步骤级）机制，减少信息孤岛、提升集体智慧。Focuses on communication (explicit symbols vs. implicit latent space) and sharing (task-level/step-level) mechanisms to reduce information silos and enhance collective intelligence.

assessment 评估体系与应用图谱：从指标到落地场景 Evaluation & Application: Metrics to Scenarios

综述给出可操作的评价维度，并梳理了记忆在不同应用中的价值与形态。

Provides actionable evaluation dimensions and maps the value of memory in applications.

多维指标Metrics：检索准确性、动态更新能力、高级认知能力（泛化/时间感知）、工程效率（延迟/Token 开销）。Retrieval Accuracy, Dynamic Update Ability, Advanced Cognitive Abilities (Generalization/Temporal Awareness), and Engineering Efficiency (Latency/Token Overhead).
应用场景Applications：对话助手（如 ChatGPT Memory）、具身机器人（Voyager）、医疗诊断、内容创作、科学发现等。Conversational Assistants (e.g., ChatGPT Memory), Embodied Robots (Voyager), Medical Diagnosis, Content Creation, Scientific Discovery, etc.

AI记忆综述 AI Memory Survey