arXiv上这篇论文解决了编码智能体记忆该存什么、什么时候用的问题。它把读写记忆的成本算进训练目标,在ChainSWE基准上让27B模型解决率从21.0%提到25.7%,还少走17%的步数。做智能体记忆或SWE agent的团队值得对比看看。
#memory mechanisms
ARC Prize 放出了 Alexis Fox 的演讲视频,讲怎么用程序化记忆让 Agent 在长任务里不掉链子,她团队还做过 ARC-AGI-3 的红队测试。做智能体的朋友可以直接抄作业,比泛泛谈记忆架构的干货多。
Anthropic 用实验数据讲清了 Agent 记忆怎么存、怎么整理才有用,61% 覆盖率和省 2 万 token 的细节很实在。
Anthropic 拆了 Claude 智能体的记忆架构,让模型在空闲时"做梦"纠错总结,CL-Bench 得分从 0.65 拉到 0.83,做智能体的可以抄作业。
Anthropic 讲了 Claude 记忆怎么做"做梦"整理,会自动纠错、加速检索,做 Agent 记忆设计的话很值得翻原文看看实现思路。
The practical implication is a hybrid memory design: keep a text store for recent and never-recorded facts, and distill stable user knowledge into model weights. Weight memory alone forgets older facts and cannot flag gaps, so the router matters. Engineers building personalised agents can test this against MemOS-style retrieval pipelines. The paper is available on arXiv (2611.05636) with a Chat with Paper interface from dair.ai.
OpenClaw 这期播客演示了 Scribes 和记忆整合,智能体记忆不再只靠聊天记录,做智能体的可以看看他们怎么存知识。
Memory benchmarks so far test only storing and recalling facts, so unlearning failures have gone unmeasured; BEMULA gives that gap its first systematic yardstick. Teams building agents that retain user data should treat delete requests as an unverified UI state and engineer an explicit delete path with adversarial extraction tests. The paper is academic work with no regional availability angle, and its closest comparison is the memory features of mainstream agent frameworks rather than another benchmark.
Splitting agent memory by problem type, rather than one shared log, delivers both better task performance and much lower token spend at the same time. Practitioners building agents that handle mixed task types can apply this by routing each new problem to a memory pool of similar past attempts. The 65% to 72% API cost reduction makes it especially relevant for anyone running agents at volume on metered model APIs.
一篇记忆机制的论文,把检索改成局部遍历,在 MiniGrid Memory 上用五 token 上下文就跑出 0.95 回报,训练步数比绝对编址少好几倍。
Databricks 拿自家 114 个任务做测试,加记忆后智能体成功率翻了三倍,文中讲清了情景记忆和派生记忆怎么用。
一个叫 MemT 1.0 的开源模型,能记住你昨天的反馈并用在今天类似任务上,靠的是 MoS 多存储专家架构,和普通 MoE 不一样。
这篇论文提出了把记忆直接做进模型计算里的 MemT-1.0 架构,128K 长度下推理快 1.86 倍、存储省 66 倍,做智能体的可以看看记忆方案怎么做。
Memory write-injection looks like a cheap way to add knowledge to a deployed model, but these results show the added facts cannot be chained into multi-hop reasoning by the model itself — orchestration must live in the application layer. Builders wiring RAG-style memory into agents should chain written atoms explicitly in prompts or system logic, and benchmark vendors should adopt coverage, synergy and distortion reporting. The tested beds include Chinese frontier models Qwen3 and DeepSeek, making the findings broadly relevant across model families.
把知识写进模型记忆到底行不行?这篇论文在 Qwen 和 DeepSeek 上做了实测,命中率数据挺扎眼,做记忆机制研究的值得看数据。
The dual-memory split addresses a core limitation of recurrent models, where short-term computation and long-term storage compete for a single hidden vector. Researchers working on state-tracking architectures or memory-augmented networks can compare CMM's results against CTM and Transformer baselines in the paper. Its selective use of long-term memory suggests a practical route to scaling recurrence beyond fixed context lengths.
arXiv 上的新论文,教长期智能体把过往推理经验压成一个轻量策略,实验里把 DeepSeek 的推理 F1 从 0.789 拉到 0.868,做智能体记忆的可以看看。
Anthropic 升级 Claude 记忆功能,Cowork 与聊天无缝对接,方便用户管理信息,无需重复解释。敏感信息保护到位,安全可靠。
Proteus给长上下文模型加了个记忆开关,逐步解锁容量,让SWLA、Titans这些模型在长文本上表现更好,越长的上下文提升越明显。
TEPA 这篇论文把记忆过期当成核心问题,用可撤销机制让智能体不被陈旧信息带偏。逆转测试里从 0.21 提到 0.95,效果很明显。
论文提出LiveMem,给LLM加了个永久记忆状态,上下文窗口再长也不怕忘事。在LongMemEval上跑分领先,值得一看。
论文用Qwen3-8B做实验,发现KV缓存里的事件行能独立保存计算视图,主动构造回答无关事件能让恢复率从6%跳到51%,对搞Agent记忆的人很有参考价值。
这篇论文提出一种新的动态视角合成方法,通过分开记忆更新和应用的频率,实现了实时处理,而且能记住长视频里的短暂遮挡区域。
微软研究员拿Transformer和人类比记忆:听五小时故事,它能记住新密码,你行吗?看看具体差异在哪。
这篇论文搞了个EvolveNav,让机器人自己从失败中学习,零样本导航成功率直接涨了10%,而且没走那么多冤枉路。
Fable 5 跨 session 记忆覆盖率73%
做视频生成或世界模型研究的团队,这篇论文帮你拆解了记忆机制中容量、压缩、读取和循环四个关键维度,看完能直接指导你的模型设计。
做AI智能体开发的团队会立刻警觉——你精心设计的记忆系统可能在悄悄退化。这篇论文用实验数据戳破了「自动总结记忆」的幻觉,建议所有用LLM做长期任务的开发者点开看看,别让记忆成为瓶颈。
eMoT 解决了大模型在多步推理中容易出错和产生幻觉的痛点,做推理增强或数学推理的开发者可以直接参考其记忆腐蚀和符号锚定机制,值得一试。
做AI智能体开发的团队值得关注——记忆机制是当前瓶颈,这篇论文直接挑战了“记忆越多越好”的假设,看完会重新思考你的记忆策略。
铁锤用大白话把AI Agent记忆讲透了,做智能体开发的团队看完能少走弯路,建议直接点开感受一下。