← 返回 2026-10-22

面向AI智能体的记忆中心持续学习 Memory-Centric Continual Learning for AI Agents

Dongfang Li, Zhuoen Chen, Bonian Jia, Zixuan Liu, Zhengxuan Lu, Can Deng, Yukun Shi, Yuhang Meng, Xin Zhang, Jiahe Huang, Fuhao Li, Junmai Wang, Shunyao Wen, Jingjun Sun, Jiaxu Tian, Jiangling Zhang, Zihan Wang, Liangyou Li, Yitong Yao, Shuangyong Song, Beilun Wang, Pengxu Wei, Hanlin Xu, Chang Liu, Yongxiang Li, Baotian Hu, Xiangyang Ji, Min Zhang 📅 2026-10-21 👍 22 2026-11-11 18:30
LLM Agent 持续学习 智能体记忆 综述/立场论文 记忆缩放律 评估方法

提出记忆中心持续学习框架:三轴分类、双环生命周期、行为化评估与记忆缩放律

前置知识

RAG(检索增强生成)与外部记忆缓存

RAG 先用向量检索等手段从外部知识库中找到与查询相关的文本片段,再把这些片段拼进 prompt 让 LLM 参考生成。多数智能体记忆系统本质上是把交互历史写入向量库/图数据库,推理时检索 Top-$k$ 条记录注入上下文,属于'检索-回填'范式。

本文的核心批评正是:现有记忆系统大多只是这种被动检索缓存,评测也只测'能否召回'。理解 RAG 的工作方式,才能理解作者为什么主张记忆应以'能否改善未来行为'而非'检索了什么'来评判。

灾难性遗忘与稳定性-可塑性困境

神经网络在顺序任务上连续梯度更新时,新任务会覆盖旧任务的权重,导致旧能力骤降,即灾难性遗忘。稳定性-可塑性困境指:要学新知识需要参数可塑,要保留旧知识需要参数稳定,二者天然冲突,是持续学习几十年的核心矛盾。

本文的中心论点是:基于基础模型的智能体不必依赖梯度更新做持续学习,记忆系统本身可承载学习,但这只会把两难困境'搬家'而非消除——外部记忆中也有功能性遗忘、检索竞争和记忆污染。

参数化与非参数化记忆

非参数化记忆指经验存放在模型权重之外(文本日志、向量索引、知识图谱),随取随用;参数化记忆指把经验写进权重、LoRA 适配器或记忆模块参数。二者的读写成本、可编辑性、容量上限完全不同:前者写快读贵可回滚,后者写贵读快难删除。

本文缩放律的一半篇幅用于论证两种记忆共享幂律前沿,但参数化记忆多出一个由载荷比 $\rho$ 驱动的干扰项,且读预算固定、写入不可逆。不懂这个对比就看不懂第 4.3 节的核心贡献。

缩放律(Scaling Law)与幂律拟合

大模型领域的缩放律指损失随参数量、数据量、算力按幂律 $L \propto x^{-\alpha}$ 下降的经验规律。本文把这一思想迁移到记忆系统:性能随累积经验 $E$、存储容量、读写预算的增长也服从资源受限的幂律前沿,超出临界载荷比 $\rho^\star$ 后干扰项接管。

记忆缩放律是本文最重要的理论产出,回答'经验、容量、算力应如何协同扩展'。理解幂律形式和'瓶颈资源主导误差'的 max-近似结构,才能读懂公式 (4.2) 与 (4.8) 及其可检验预测。

强化学习中的信用分配(Credit Assignment)

信用分配指当奖励延迟到达时,如何确定序列中哪些早期动作对最终结果负责。智能体在第 $t$ 步决定'写入/删除某条记忆',其真实价值可能在 $t+100$ 步之后才显现,且中间经过多轮摘要、剪枝与任务切换,导致训练记忆管理策略 $\pi_{mem}$ 极为困难。

本文外环(策略演化环)能否闭环的关键障碍就是延迟信用分配。第 6.1 节将其列为第一大开放挑战,并指出路径依赖(写/删改变后续环境)会破坏标准 RL 假设。

研究动机

基础模型驱动的智能体已具备多步推理、规划和工具调用能力,但无法在长期部署中从经验中可靠改进。现代智能体框架虽能系统记录执行日志、时间戳和环境状态,却把积累的历史当作待检索的原始文本,检索后直接塞回 prompt——而随交互轨迹增长,未管理的上下文会稀释相关信号、增加推理延迟、退化注意力。两条传统路线各有硬伤:扩大上下文窗口受会话长度限制且长上下文中性能随序列长度、证据位置和任务复杂度急剧下降;持续微调则需要大量计算、面临灾难性遗忘且缺乏可靠的错误回滚机制。现有智能体记忆研究虽有快速进展,但多数聚焦于具体工程实现:采用信息检索视角,把记忆当作外部数据库或 RAG 缓存;已有分类法混淆了'经验存在哪''以什么形式存在''操作如何管理'这三个不同的系统选择。评测同样局限——LoCoMo、LongMemEval 等主流基准大多给智能体一段静态历史再问答,测的是召回而非'部署期积累的经验能否在干扰与漂移下改善后续行动'。更糟的是,随存储增长,未管理的外部记忆会出现检索噪声、记忆污染和任务分布漂移下的负迁移,导致智能体反复犯同样的错误、无法处理冲突或过时信息。

本文的目标是本文提出记忆中心持续学习(Memory-Centric Continual Learning, MCCL)框架,目标是把智能体记忆从被动存储组件重构为主动的学习基础设施,回答一个根本问题:记忆系统能否在不改变基础模型权重的前提下,让智能体在部署中持续吸收反馈、修正策略、解决知识冲突。具体目标分解为四项贡献:其一,建立统一的三轴记忆分类法(载体 substrates / 资产 assets / 策略 policies),把物理存储、认知表示与操作控制解耦;其二,将记忆动态形式化为五阶段生命周期(选择性形成、结构化巩固、再激活利用、知识修订与选择性遗忘、反馈驱动的策略演化),构成内环记忆状态环与外环策略适应环的耦合动力系统,通向递归自我改进(RSI);其三,建立行为级评估框架,用四门保留级联和 stateful-reset 匹配对照把'真正的持续学习增益'与'单纯上下文扩展'区分开;其四,提出记忆缩放律,建模智能体性能如何随经验、存储容量与计算预算在非参数与参数记忆两种载体上缩放,将研究焦点从'更大的记忆是否有用'转向'经验、容量与算力应如何协同扩展'。

与已有工作不同的是,本文的独特切入角度体现在四个方面。第一,评判标准的反转:作者主张记忆系统应以其'能否改善未来行动'来评价,而非'检索出什么'——这直接区别于以召回率、检索精度为中心的既有评测传统,并因此提出四门级联(preservation→access→utilization→effect)来定位失败点。第二,三轴解耦:已有综述常把记忆按内部/外部或工程实现分类,本文坚持按状态物理持久化的位置(substrate)、可复用表示的形式(asset)、管理决策的控制方式(policy)三个正交轴划分,澄清了'拥有外部存储不等于拥有可迁移能力'。第三,对稳定性-可塑性困境的重新定位:作者论证该困境不必然依赖连续梯度更新,记忆本身可作为演化的学习基础设施,但外部存储只是把困境迁移到检索与维护层面而非消除它。第四,方法论缺口:持续学习文献聚焦权重更新,记忆文献聚焦工程实现,两社区缺乏共同语言——本文用耦合动力系统 $M_{t+1}=T(M_t,e_t;\pi_{mem}^t)$ 与 $\pi_{mem}^{t+1}=U(\pi_{mem}^t,F_{t:t+H})$ 把二者统一,并首次把缩放律框架同时延伸到非参数与参数记忆,指出参数记忆独有的载荷干扰项 $I_{s,\pi}(\rho)$ 是 substrate 的内生性质而非检索问题。

核心方法

MCCL 把持续学习建模为一个耦合动力系统,直觉上就是'记忆内容在变'与'管理记忆的规则也在变'两个嵌套循环。技术路线分四步展开。第一步是架构分类:三轴分类法把设计空间拆开——载体(WHERE)决定经验在什么物理介质上持久化,分为上下文记忆(prompt 内 token)、激活记忆(KV 缓存、隐状态、记忆槽)、外部可寻址记忆(文本日志、向量索引、知识图谱、层级树)与自适应参数记忆(可更新的权重/适配器/快速参数);资产(WHAT)按检索时的用途分为情节(episodic,保留事件与情境)、语义(semantic,抽象事实与关系)与程序性(procedural,工作流、技能、可执行代码);策略(HOW)分为规定式(FIFO/LRU/显著性打分等固定规则)、审议式(LLM 逐案例决定写/取/改)与学习式(从结果反馈中训练记忆动作控制器)。第二步是生命周期引擎:以 $M_{t+1}=T(M_t,e_t;\pi_{mem}^t)$ 描述内环,外环以 $\pi_{mem}^{t+1}=U(\pi_{mem}^t,F_{t:t+H})$ 描述策略演化,且只有当记忆改变引发下游行为效应 $\Delta B_{t+k}$ 时才算学习发生。第三步是行为化评估:保留、适应、迁移三个结果维度加四门级联与 $\Delta P_t$ 匹配对照。第四步是缩放律:非参数版本 $\epsilon^\star-\epsilon_\infty\asymp\max_j a_j x_j^{-\alpha_j}$($x$ 含经验、容量、读预算、写预算),参数版本额外叠加干扰项 $I_{s,\pi}(\rho_t)$,其中 $\rho_t=I(E_t)/(\kappa M)$、$\kappa\approx2$ 比特/参数。

核心创新点有四,均与既有工作存在本质区别。其一,'记忆即学习基础设施'的立场转换:传统持续学习把学习等同于权重更新,本文论证记忆状态本身的演化就是一种持续学习,前提是它满足'保留-访问-利用-行为效应'四门级联——这给出了'weakest-gate 假设':行为保留由最不可靠的一环瓶颈化,增加容量可能提高 preservation 却因检索竞争加剧反而降低 access。其二,$\Delta P_t$ 匹配对照协议:$\Delta P_t=P_{stateful}(t)-P_{reset}(t)$ 本身不是新指标,但本文系统化了它的使用条件——必须同时控制无记忆基线、全历史上下文基线、静态 RAG 基线和等计算基线,并把 reset 的含义按 substrate 分别定义(外部库可做 write-off/read-off/clear-store 干预,参数记忆则需区分在线更新与读出)。其三,参数记忆缩放律中的载荷比干扰项:所有写入共享同一组参数,新写入会改变旧内容的读出方式,故 $\epsilon^\star-\epsilon_\infty\asymp\max_j a_j x_j^{-\alpha_j}+I_{s,\pi}(\rho_t)$,干扰在载荷比低于 $\rho^\star_{s,\pi}$ 时近零、越界后上升,且 $\rho^\star$ 随载体不同可差数量级(论文示例中 backbone editing $\rho^\star=3$、专用模块 $\rho^\star=300$)。其四,复用经济学:$C_N=C_{build}/N+C_{maint}+C_{query}$ 把'构建成本随复用次数摊销'形式化,说明保留、适应、迁移等持续学习性质如何决定记忆策略的划算程度。

方法步骤详情

方法按'分类→生命周期→评估→缩放律'四步执行。第一步,给目标系统在三轴上定位:新经验先进入哪个载体(如上下文 token)、以何种资产形态持久化(如先情节后语义)、写/取/改/忘由哪层策略控制(规定/审议/学习)。第二步,沿五阶段生命周期设计与审计:'形成与写入'做准入决策、选表示结构与粒度、记录来源/时间范围/置信度以防记忆污染;'组织与巩固'决定经验单元边界(动态事件检测)、巩固时机(LycheeMemory V2 以语义段边界触发、RecMem 以复现信号触发)和抽象层级,并保留指向原始情节的证据链接以防上下文坍缩;'检索与利用'要求恢复足够支撑证据、按当前决策状态匹配记忆而非仅按查询相似度,并区分'被检索到'与'被实际使用';'修订与遗忘'区分冲突解决、持久化修订与三种遗忘操作(硬删除、按窗口过期、软抑制),事实有效性与检索可用性是独立属性;'策略演化'区分状态级写入、离线控制器训练(AgeMem/UMA 用 RL 训练后冻结)与在线策略演化三层。第三步,按评估蓝图实施:定义能力锚点并采集前后测;在声明的延迟-干扰网格上设检查点;运行 stateful 与 reset 双条件(骨干、工具、上下文、检索与推理预算全部对齐);填写重测矩阵 $R_{t,i}$(对角测获取、下三角测保留、held-out 测迁移),并维护记录时间、写入、冲突与成本的事件账本。第四步,用缩放律做资源规划:扫描经验 $E$、容量、读/写预算,识别瓶颈项,按 $\bar{B}_0\propto(E/E_0)^{\alpha_E/\alpha_B}$ 平衡扩展;参数记忆还需保证 $\rho<\rho^\star$,即 $M_{req}\gtrsim I(E)/(\kappa\rho^\star)\propto E^\nu$。

技术新颖性

技术新颖性体现在四个层面。概念层面:本文是首批明确拒绝'检索缓存'范式、把智能体记忆当作持续学习载体的系统性立场工作,与此前以工程实现组织内容的记忆综述不同,它提供了可诊断、可组合的设计语言(同一系统可在三轴上做不同组合,例如'情节资产+外部可寻址载体+审议策略')。评测层面:四门级联把'记住 $x$'(preservation+access)与'从 $x$ 中保留了学习'(还需 utilization+effect)在操作上分开,weakest-gate 假设可直接指导诊断——例如提升容量反而可能因检索竞争伤及 access 门;同时它把基准按'可支持的学习主张强度'分层(Table 2):静态历史问答只支持'记住并使用',顺序交互环境加 stateful-reset 对照才接近持续学习主张。理论层面:记忆缩放律是全新尝试——经典缩放律针对预训练损失,本文将其扩展到部署期经验积累,且识别出参数记忆与非参数记忆的三点本质差异(读预算由架构固定、存在载荷比干扰项、容量需求被 $\rho^\star$ 放大),并提出第三个可检验但尚未被检验的预测:不同容量下测得的载荷曲线按 $\rho$ 归一后应坍缩为单一曲线。证据组织层面:论文没有只做理论推演,而是用自做实验(Mem0 与 Raw RAG 在 LoCoMo/LongMemEval/IdeaBench 上)验证非参数侧的存储-读取互补与无巩固时经验累积的负增益,并系统汇编参数记忆文献数字(Memory Layers、LoRA 记忆化、ROME/MEMIT 编辑寿命、GradMem 写预算实验、TMEM 写入时机)来支撑干扰项的存在性与 $\rho^\star$ 的载体依赖性。

Memory scaling across two substrates. Theoretical illustrations: left, joint allocation of storage and reading resources sustains improvement beyond fixed-capacity limits; center, greater experience requires more parametric memory capacity; right, load-induced interference emerges at substrate-dependent thresholds. Curves are synthetic illustrations; see Section 4.3 and Figures 7 and 11.
Figure 1: Memory scaling across two substrates. Theoretical illustrations: left, joint allocation of storage and reading resources sustains improvement beyond fixed-capacity limits; center, greater experience requires more parametric memory capacity; right, load-induced interference emerges at substrate-dependent thresholds. Curves are synthetic illustrations; see Section 4.3 and Figures 7 and 11.
Memory-centric continual learning as a coupled dynamical system. A task stream is executed by the agent and produces outcome feedback. Experience is organized as episodic, semantic, and procedural assets that persist across contextual, activation, external, and parametric substrates. An inner memory-state loop forms, consolidates, reactivates, and revises these assets, while prescribed, deliberative, or learned policies govern the corresponding operations. An outer policy-adaptation loop uses task feedback to update memory management.
Figure 2: Memory-centric continual learning as a coupled dynamical system. A task stream is executed by the agent and produces outcome feedback. Experience is organized as episodic, semantic, and procedural assets that persist across contextual, activation, external, and parametric substrates. An inner memory-state loop forms, consolidates, reactivates, and revises these assets, while prescribed, deliberative, or learned policies govern the corresponding operations. An outer policy-adaptation loop uses task feedback to update memory management.
The memory lifecycle as two coupled processes in memory-centric continual learning. The memory-state loop governs how experience enters persistent memory, is organized and reused, and is revised or forgotten as conditions change. The control loop uses behavioral outcomes to influence future memory decisions.
Figure 4: The memory lifecycle as two coupled processes in memory-centric continual learning. The memory-state loop governs how experience enters persistent memory, is organized and reused, and is revised or forgotten as conditions change. The control loop uses behavioral outcomes to influence future memory decisions.
Lifecycle–control landscape of selected representative agent memory systems. The horizontal axis shows the memory operations emphasized by each system. The vertical axis shows how decisions are controlled: through prescribed rules, deliberative runtime LLM reasoning, or learned control. Darker regions indicate the main design focus and lighter regions indicate supporting operations. Colors distinguish design families. Positions describe design choices, not performance.
Figure 5: Lifecycle–control landscape of selected representative agent memory systems. The horizontal axis shows the memory operations emphasized by each system. The vertical axis shows how decisions are controlled: through prescribed rules, deliberative runtime LLM reasoning, or learned control. Darker regions indicate the main design focus and lighter regions indicate supporting operations. Colors distinguish design families. Positions describe design choices, not performance.
Matched comparison framework for isolating the contribution of agent memory. The stateful condition retains accumulated experience, whereas the reset condition removes experience-related state on a matched task. $\Delta P_t = P_{stateful}(t) - P_{reset}(t)$ is the additional benefit associated with accumulated state. No-memory, full-history, static RAG, and matched-compute baselines progressively rule out alternative explanations.
Figure 6: Matched comparison framework for isolating the contribution of agent memory. The stateful condition retains accumulated experience, whereas the reset condition removes experience-related state on a matched task. $\Delta P_t = P_{stateful}(t) - P_{reset}(t)$ is the additional benefit associated with accumulated state. No-memory, full-history, static RAG, and matched-compute baselines progressively rule out alternative explanations.
Theoretical illustration of joint scaling and optimal resource allocation. (a) Fixed-capacity error curves and their optimized lower envelope. (b) Returns at equal total budgets. (c) Optimal capacity and reading budgets. (d) Equal-error contours and the optimal path.
Figure 7: Theoretical illustration of joint scaling and optimal resource allocation. (a) Fixed-capacity error curves and their optimized lower envelope. (b) Returns at equal total budgets. (c) Optimal capacity and reading budgets. (d) Equal-error contours and the optimal path.
Theoretical illustration of experience structure, capacity requirements, and adaptation. Curves use preset response and recovery functions. (a) Performance as task demand expands with experience, at fixed capacity limits $\bar{B}$. (b) Capacity needed to maintain a target under different experience structures. (c) Recovery after environmental changes, marked by dashed lines; $m$ is the maintenance rate.
Figure 9: Theoretical illustration of experience structure, capacity requirements, and adaptation. Curves use preset response and recovery functions. (a) Performance as task demand expands with experience, at fixed capacity limits $\bar{B}$. (b) Capacity needed to maintain a target under different experience structures. (c) Recovery after environmental changes, marked by dashed lines; $m$ is the maintenance rate.
Theoretical illustration of reuse and resource-optimal policy choice. (a) Query-cost–performance Pareto frontiers; lighter points denote dominated configurations. (b) Amortization at a common performance target $P=0.75$. (c) The best policy at each reuse count and average-cost budget; gray denotes infeasibility, and dashed lines are performance contours.
Figure 10: Theoretical illustration of reuse and resource-optimal policy choice. (a) Query-cost–performance Pareto frontiers; lighter points denote dominated configurations. (b) Amortization at a common performance target $P=0.75$. (c) The best policy at each reuse count and average-cost budget; gray denotes infeasibility, and dashed lines are performance contours.
Theoretical illustration of the parametric memory scaling relation. (a) Excess loss versus memory capacity at different experience volumes; interference dominates until the load ratio falls below $\rho^\star$, after which the capacity term follows a power law. (b) Excess loss versus accumulated experience at fixed capacity for backbone editing (low $\rho^\star$) and a dedicated memory module (high $\rho^\star$); dotted lines mark the critical loads. (c) Capacity required to maintain a target performance for independent facts ($\nu=1$) and shared structure ($\nu<1$).
Figure 11: Theoretical illustration of the parametric memory scaling relation. (a) Excess loss versus memory capacity at different experience volumes; interference dominates until the load ratio falls below $\rho^\star$, after which the capacity term follows a power law. (b) Excess loss versus accumulated experience at fixed capacity for backbone editing (low $\rho^\star$) and a dedicated memory module (high $\rho^\star$); dotted lines mark the critical loads. (c) Capacity required to maintain a target performance for independent facts ($\nu=1$) and shared structure ($\nu<1$).

实验结果

自做实验统一用 Qwen3-30B-A3B-Instruct-2507 做构建/答题/打分、qwen3-embedding-0.6b(1024 维)检索,对比 Mem0 0.1.94 与 Raw RAG。实验一(LoCoMo 容量×读取扫描,10 对话 1540 问题、4×4 网格):读取预算 512→2048 tokens 时,Raw RAG F1 在 1/8 历史保留仅 +0.89 pp、全历史 +3.82 pp,配对差 2.93 pp(95% CI [2.06, 4.02])——证实存储与读取互补。实验二(LongMemEval 100 样本):同口径准确率 35.00%→44.00%。实验三(IdeaBench 顺序部署,50 固定查询、3 随机顺序):Mem0 默认在线模式的 $\Delta P_t$ 非单调——8192-token 容量下从 64 训练对话后的 +7.10 pp 跌至 128 对话的 −10.43 pp、256 对话的 −13.93 pp;占用平台期 3,858 tokens(8192/32768 上限均未触发);每 16/64 段巩固可把占用降到约 1800 tokens 但最终增益仍为负——无巩固的持续写入让经验从资产变成负担。实验四(LoCoMo 查询经济学,9 对话 180 问题):Read all F1 4.63%(729.78 tokens、0.154s)、Search 4.29%(225.68、0.154s)、Search+summarize 3.91%(43.52、0.367s);摘要比全库读取配对低 0.72 pp([-1.14, -0.27])但省 16 倍输入、慢 2.4 倍。参数记忆侧汇编文献证据:Memory Layers 把 1.3B 配置的 Memory+ 值 1M→64M,NaturalQuestions 13.68→20.78、TriviaQA F1 42.89→62.14(dense 无记忆层仅 7.76/32.64),但 16M→64M 区间 NQ 仅 +0.64 而 TriviaQA +3.47;LoRA 记忆化服从 $\Delta L=Cr^{\alpha}\ell^{-\beta}+b$($R^2>0.98$),rank 128 下 SFT 40.0%、MemFT-OT 100.0%、MemFT-SW 94.4%;GradMem 显示写预算可替代容量(固定 8 向量,5 写步 96 对 88.4% vs 1 写步 32.6%);顺序 ROME 约 100-1000 次退化、顺序 MEMIT 约 1400 次灾难遗忘而 batch 一次可 10000 条,WISE 达 3000 次、MEMORYLLM(约 1.07B 池)650000 次更新稳定——两者 $\rho^\star$ 差一到两个数量级;TMEM 写入阈值倒 U 形:LoCoMo 4K 峰值 F1 25.72、LongMemEval-S 12K 峰值 41.24。

Diagnostic scope of evaluation metric layers in agent memory systems. While component-level and snapshot metrics diagnose internal storage fidelity and retrieval reachability, establishing continual learning requires behavioral retention, adaptive revision, and cross-task transfer.
Table 1: Diagnostic scope of evaluation metric layers in agent memory systems. While component-level and snapshot metrics diagnose internal storage fidelity and retrieval reachability, establishing continual learning requires behavioral retention, adaptive revision, and cross-task transfer.
Categorization of representative agent memory benchmarks by evaluation paradigm. Benchmarks are organized by the extent to which their experimental design supports genuine learning-from-experience claims rather than static information retrieval.
Table 2: Categorization of representative agent memory benchmarks by evaluation paradigm. Benchmarks are organized by the extent to which their experimental design supports genuine learning-from-experience claims rather than static information retrieval.
Experimental comparisons in this section. Repeated predictions across budgets are aggre- gated within the stated sampling units. LoCoMo scores use categories 1–4.
Table 3: Experimental comparisons in this section. Repeated predictions across budgets are aggre- gated within the stated sampling units. LoCoMo scores use categories 1–4.
Three reading policies over Mem0 memory banks on nine LoCoMo conversations and 180 questions, with three timing repetitions. Memory input includes serialization and formatting. Median latency covers retrieval, query-time summarization, and answering, excluding offline construction. F1 intervals use 10,000 conversation-cluster bootstrap resamples.
Table 4: Three reading policies over Mem0 memory banks on nine LoCoMo conversations and 180 questions, with three timing repetitions. Memory input includes serialization and formatting. Median latency covers retrieval, query-time summarization, and answering, excluding offline construction. F1 intervals use 10,000 conversation-cluster bootstrap resamples.
Measured performance across capacity, reading, and experience settings. (a–b) Stored text and Raw RAG performance on LoCoMo and LongMemEval. (c) Paired gains from increasing LoCoMo reading from 512 to 2,048 tokens; error bars in (a–c) are cluster-bootstrap 95% intervals. (d) Stateful–reset score differences on fixed IdeaBench questions at successive checkpoints, without periodic consolidation; shading spans the min–max range across three random training-dialogue orderings.
Figure 8: Measured performance across capacity, reading, and experience settings. (a–b) Stored text and Raw RAG performance on LoCoMo and LongMemEval. (c) Paired gains from increasing LoCoMo reading from 512 to 2,048 tokens; error bars in (a–c) are cluster-bootstrap 95% intervals. (d) Stateful–reset score differences on fixed IdeaBench questions at successive checkpoints, without periodic consolidation; shading spans the min–max range across three random training-dialogue orderings.
Reported capacity and load effects in published parametric-memory studies. (a) Memory+ value count at a fixed 1.3B configuration; dotted lines show the dense model without memory layer. (b) Memorization accuracy versus LoRA rank for three write methods on Qwen3-8B. (c) Reported exact match versus stored key–value pairs with eight fixed memory vectors and different numbers of write steps. (d) Reported sequential or batch update horizons across parametric-memory studies.
Figure 12: Reported capacity and load effects in published parametric-memory studies. (a) Memory+ value count at a fixed 1.3B configuration; dotted lines show the dense model without memory layer. (b) Memorization accuracy versus LoRA rank for three write methods on Qwen3-8B. (c) Reported exact match versus stored key–value pairs with eight fixed memory vectors and different numbers of write steps. (d) Reported sequential or batch update horizons across parametric-memory studies.
Reported effects of the foundation model and write timing on parametric memory. (a) Average domain perplexity (biomedicine, finance, and law) of Qwen2.5 models from 0.5B to 72B, with and without the same 0.5B Memory Decoder; arrows mark the reduction at the smallest and largest models, and the dotted line marks the 72B model alone. (b–c) F1 and exact match of TMEM on Qwen3-4B as the context threshold that triggers a parametric write varies, on LoCoMo and LongMemEval-S; hollow markers denote the observed maxima.
Figure 13: Reported effects of the foundation model and write timing on parametric memory. (a) Average domain perplexity (biomedicine, finance, and law) of Qwen2.5 models from 0.5B to 72B, with and without the same 0.5B Memory Decoder; arrows mark the reduction at the smallest and largest models, and the dotted line marks the 72B model alone. (b–c) F1 and exact match of TMEM on Qwen3-4B as the context threshold that triggers a parametric write varies, on LoCoMo and LongMemEval-S; hollow markers denote the observed maxima.
查看结构化数据
任务指标本文基线提升
LoCoMo 历史问答(存储×读取联合扫描) 词面 F1(pp),cluster-bootstrap 95% 区间 读取预算 512→2048 tokens:全历史保留 +3.82 pp;1/8 历史保留 +0.89 pp 读取预算 512 tokens(同一存储水平) +3.82 pp(全历史),配对差 2.93 pp,95% CI [2.06, 4.02];证实存储-读取互补
LongMemEval 语义问答(容量扫描) 本地语义裁判准确率(%) 读取 512→2048 tokens 时全历史 Raw RAG 从 35.00% 升至 44.00% 读取预算 512 tokens +9.00 pp,与 LoCoMo 一致支持存储-读取互补性
IdeaBench(MemoryBench)顺序部署,50 固定目标查询 stateful-reset 性能差 $\Delta P_t$(pp),3 种任务顺序 Mem0 默认模式:+7.10 pp(64 训练对话)→ −10.43 pp(128)→ −13.93 pp(256);无巩固时占用平台 3,858 tokens;每 16/64 段巩固后占用约 1,800 tokens 但最终增益仍为负 同任务 reset 条件(无累积经验) 无提升反而恶化:无巩固的持续写入使经验累积从正增益转为 −13.93 pp,凸显记忆巩固与遗忘的必要性
LoCoMo 查询经济学(9 对话、180 问题) F1(%)、序列化输入 tokens、端到端延迟(s) Read all:F1 4.63 / 729.78 tokens / 0.154s;Search:4.29 / 225.68 / 0.154s;Search+summarize:3.91 / 43.52 / 0.367s 三种读取策略互为对照 查询时摘要省 16 倍输入但 F1 配对下降 0.72 pp(CI [-1.14, -0.27])且延迟增至 2.4 倍——策略选择取决于哪个资源受限
参数记忆容量扩展(文献汇编:Memory Layers at Scale,1.3B 配置) NaturalQuestions 准确率 / TriviaQA F1 Memory+ 值 1M→64M:NQ 13.68→20.78,TriviaQA 42.89→62.14(128B 记忆参数可配 ≤8B 底座) 同配置无记忆层的 dense 模型(NQ 7.76 / TriviaQA 32.64) 大幅超越 dense 基线;但 16M→64M 区间 NQ 仅 +0.64 而 TriviaQA +3.47,支持任务相关的幂律前沿
参数记忆写入方法与时机(文献汇编) 记忆化准确率(%,rank 128,Qwen3-8B)/ F1(TMEM) MemFT-OT 100.0% vs MemFT-SW 94.4% vs SFT 40.0%;TMEM 写入阈值 LoCoMo 4K 峰值 F1 25.72、LongMemEval-S 12K 峰值 41.24(倒 U 形) 同 rank 下 SFT 写入 / 其他阈值 写入方法改变缩放律系数 $a_M$ 与 $\alpha_M$;写入时机存在最优区间,过频或过稀均退化

局限与改进

作者明确承认的限制包括:缩放律中的指数 $\alpha_j$、$\nu$ 与临界载荷比 $\rho^\star$ 均为经验量,论文只给出函数形式与合成插图(如 $\epsilon_{ex}(b,r)=0.6b^{-0.55}+0.4r^{-0.75}$ 的示例参数),未对任何真实系统做系统性拟合;干扰项的幂律形式仅从 LoRA 微调遗忘的移位幂律间接推出,第三个可检验预测——不同容量下的载荷曲线按 $\rho$ 归一后应坍缩为单一曲线——尚无实验检验;参数记忆的全部证据取自他人论文的既有数字('taken from the original papers without fitting'),且各研究协议互不兼容,只能比较数量级。我自己的观察补充四点:第一,自做实验规模很小(LoCoMo 10 段对话 1540 问题、LongMemEval 100 样本、IdeaBench 50 条查询),且只覆盖 Mem0 与 Raw RAG 两个系统,未涉及 A-MEM、Zep、MemGPT 等更新型系统,缩放律的普适性存疑;第二,Table 4 中 F1 绝对值极低(4.63%),说明 LoCoMo 类目 1-4 对这些朴素读法本身极难,结论的方向性可靠但外推到能力更强的记忆系统时未必保持;第三,IdeaBench 的负增益结论可能被 Mem0 默认在线写入污染(论文自述 $t=0$ 时有非零偏移,评估期间也在写入),−13.93 pp 的归因需谨慎;第四,外环'在线策略演化'只有概念框架(式 3.2),论文没有给出任何具体算法或损失函数实例,距可实施还有相当距离。

独立分析的弱点

第一,实证薄弱与理论抱负不匹配:缩放律是全文最大卖点,但自做实验只有 4 组小规模扫描,从未拟合出任何一个系统的 $\alpha_j$ 或 $\rho^\star$ 数值,理论曲线全部是合成插图;改进方向是在 3-5 个开源记忆系统(Mem0、A-MEM、Zep、MemGPT)上系统扫描经验/容量/读/写四资源,实际估计指数并检验载荷坍缩预测。第二,系统覆盖窄且版本单一:只测了 Mem0 0.1.94 和 Raw RAG,而 Mem0 默认在线模式在评估期仍写入,导致 $\Delta P_t$ 从起点就带偏移;改进方向是固定'评估期冻结写入'的协议重跑,或改用支持训练-评估分离的模式。第三,外环策略演化缺乏可操作算法:式 (3.2) 的更新规则 $U$ 没有具体形式,第 6.1 节只罗列了价值模型、消融日志、检查点分支等方向性想法;改进方向是把记忆动作(write/consolidate/retrieve/revise/forget)显式建模为 POMDP,用后见批评家(hindsight critic)和从同一记忆检查点分支的 rollout 解决路径依赖。第四,评估蓝图成本过高:$R_{t,i}$ 重测矩阵加事件账本加多基线对照在真实部署中开销巨大,论文未讨论近似的低成本的替代协议。第五,四门级联的'利用'与'效应'测量依赖模型自省或人工标注,缺乏自动化工具;改进方向是结合注意力归因、答案消融(去掉该记忆后性能变化)等自动探针量化 uptake。第六,可信治理一节(注入防御、回滚、删除验证)只有原则清单,没有任何基准或工具原型,是该领域公认的短板。

未来方向

作者在第 6 节提出四大挑战。其一,经验价值预测与长期信用分配:需要写前价值模型(在任务流上学习、用缩放曲线判断边际收益)、错误归因机制(写入决策日志、消融测试、摘要到原始情节的链接)与应对路径依赖的方法(从相同记忆检查点分支 rollout、后见批评家),终极目标是闭环控制——反馈不仅更新记忆内容,还更新管理方式。其二,动态环境下的稳定性-可塑性:需要写入门控过滤杂波、以原始证据校验的巩固、匹配经验适用条件的检索;评测应跟踪保留率、检索排名与实际使用率对冻结记忆基线的变化。其三,经验记忆与参数学习的双向协同:借鉴互补学习系统(CLS)、再固化(reconsolidation)与突触标签-捕获理论,构建三步循环——按新颖性/效用/频率筛选痕迹、经重放与抽象把反复验证的记忆经模块化适配器固化入权重、用新事实检验并对过时知识'重新打开'修订。其四,可信记忆治理:写入前验证来源与权限、借鉴数据库恢复与信念修正防止摘要丢源、共享库角色访问控制与工具沙箱、用压力测试与恢复目标证明删除数据及衍生残留真正清除。基于本文成果可延伸的方向还包括:把 $\rho^\star$ 测量标准化为参数记忆系统的公开指标;设计干预式记忆基准取代静态历史问答;探索外部与参数记忆的最优分工边界(快速变化经验留外部、稳定高频经验入参数);研究记忆污染的传播动力学与可回滚的分级巩固机制。

复现评估

作为综述/立场论文,本文没有发布代码仓库,但可复现性信息总体充分。自做实验部分交代了完整配置:Mem0 0.1.94(开源可装)、构建/答题/打分统一用 Qwen3-30B-A3B-Instruct-2507、检索用 qwen3-embedding-0.6b(1024 维)、token 计数用 o200k_base、指标用宏平均加会话级 cluster-bootstrap 95% 区间(10000 次重采样)、LongMemEval 用本地语义裁判、MemoryBench 用本地 rubric 打分;数据集全部公开(LoCoMo 10 段对话 1540 问题、LongMemEval 100 选取样本、IdeaBench 50 固定查询、查询经济学 9 对话 180 问题 3 次时序重复)。复现 Table 4 与 Figure 8 所需算力不大——主要成本是一个 30B 级模型的推理,用 API 替代即可,单卡或 API 预算数小时内可完成。参数记忆部分的图表(Figure 12、13)是文献数字汇编,各数字可对照原论文逐一核验,但协议互不兼容,论文自己也声明这些值未经拟合、只可比较数量级。难度评估:复现非参数实验为中等(需要正确复刻 Mem0 的默认在线配置与 capacity 上限设置,论文承认 $t=0$ 有非零偏移这类细节坑);复现或延伸缩放律拟合为高难度(需自行设计多系统多资源扫描并处理载荷坍缩检验这一前人未做的实验);复现理论公式本身无需任何实验。总体而言,这是一篇'观点可讨论、实验可复跑、理论待检验'的工作。

相似论文推荐