AI 终于学会「做梦」了：Anthropic Dreaming 如何让 Agent 在你睡觉时自我进化
AI Learns to Dream: How Anthropic's Dreaming Feature Lets Agents Self-Improve While You Sleep
> 📌 TL;DR
> Anthropic 于 2026 年 5 月 6 日发布了 Dreaming 功能，让 Claude Agent 能在任务间隙「做梦」——回顾过去的操作、发现重复错误、把经验写入长期记忆。法律 AI 公司 Harvey 任务完成率提升 6 倍，医疗文档公司 Wisedocs 处理时间减半。模型权重不变，变的是记忆质量。这可能是 AI Agent 从「工具」进化为「同事」的关键一步。
一句话说清楚：AI Agent 的最大痛点不是不够聪明，是「记不住」
你有没有这种经历：给 AI 助手布置一个任务，它搞砸了。你纠正它，它学会了。第二天再来——完全忘了，同样的错误再犯一遍。
这不是 AI 笨。这是 AI Agent 架构的根本缺陷：每次会话都是一张白纸。
2026 年 5 月 6 日，Anthropic 在旧金山举办的第二届 Code with Claude 开发者大会上，给出了一个优雅的解决方案——Dreaming（做梦）。
Dreaming 到底是什么？
简单说：让 AI Agent 在「下班后」自动复盘。
技术上，Dreaming 是一个计划调度的后台进程，它会：
1. 回放——检查 Agent 最近完成的所有会话记录
2. 分析——识别重复出现的错误模式、团队偏好、收敛的工作流
3. 整理——合并重复信息、清理过时条目、把高价值经验写入长期记忆
4. 输出——生成一组纯文本笔记和结构化「剧本」，供下次会话使用
Anthropic 把这比作人类大脑的海马体记忆巩固——你白天经历的事，大脑在睡眠时重新播放、筛选、存储。决定什么值得记住，什么可以丢掉。
关键技术细节
- 不改权重：Dreaming 不做梯度更新，不微调模型。它只策划 Agent 的持久记忆和上下文，让下一次会话从一个更干净、更准确的知识库开始
- 可观察可审计：所有学到的东西都以纯文本形式存储，人类可以随时查看、编辑、否决
- 运行速度快：典型的 Dreaming 运行只需几分钟，适合夜间或低峰期调度
- 可选人工审核：开发者可以让 Dreaming 自动更新记忆，也可以要求每次更新都经过人工确认
实战成绩：不是 PPT，是真金白银的提升
Harvey（法律 AI）：任务完成率 ×6
Harvey 是 Anthropic 最知名的企业客户之一，用 Claude Agent 做法律文书起草。他们遇到的问题很典型：Agent 总是忘记特定文件格式的坑和工具的变通方法，导致同样的法律起草任务反复失败。
开了 Dreaming 之后——不改代码、不换模型——Agent 自己把这些「坑」记住了。任务完成率直接提升了约 6 倍。（数据来源：Anthropic 官方发布，2026-05-06）
需要注意的是：这个 6 倍数据来自 Harvey 内部测试，尚未有独立第三方验证。但多个媒体（VentureBeat、SiliconANGLE、9to5Mac）都引用了相同数据。
Wisedocs（医疗文档）：处理时间 -50%
Wisedocs 使用的是和 Dreaming 同时发布的另一个功能——Outcomes（成果评估）。这是一个自评估循环：生成一个独立的「评分员」Agent，它只看最终输出和预设的评分标准，不接触原 Agent 的推理过程。
结果：文档审核时间直接砍半。Anthropic 内部测试显示，Outcomes 在最难的任务上提升了高达 10 个百分点的成功率，其中 .docx 文件生成提升 8.4%，.pptx 提升 10.1%。
Netflix：多 Agent 协作处理数百个构建日志
Netflix 利用同期发布的多 Agent 编排功能，让一个「领导」Agent 把大任务拆分成子任务，分配给多个专家 Agent——每个有自己的模型、系统提示词、工具集和独立上下文窗口。
这件事为什么重要？因为它解决了 AI Agent 的「格局」问题
目前 AI Agent 的发展瓶颈不在智力上——GPT-5.5、Claude Mythos 已经足够聪明了。瓶颈在记忆和经验积累上。
| 维度 | 传统 Agent | 有 Dreaming 的 Agent |
|------|-----------|---------------------|
| 记忆 | 每次会话清零 | 跨会话累积学习 |
| 错误 | 同一个坑踩 N 遍 | 踩一次，记一辈子 |
| 偏好 | 每次都要重新告知 | 自动发现并记住 |
| 工作流 | 固定不变 | 随使用进化 |
| 信任度 | 需要人盯着 | 可以逐步放手 |
Anthropic CEO Dario Amodei 在大会上描述了一条路线：从单个 Agent → 多个 Agent 协作 → 整个组织级智能。他重申了一个大胆预测：2026 年将出现第一家由一个人运营的十亿美元公司。
Dreaming 是这条路线上的关键基础设施。因为你不可能把关键业务交给一个「每天早上都失忆」的员工。
安全隐患：当「记忆」变成攻击面
这里必须泼一盆冷水。
安全研究人员已经指出了一个严肃的风险：如果恶意输入能说服 Agent 把错误指令当成正确经验，Dreaming 可能会把这个「毒」巩固到长期记忆中。
这本质上是 Prompt Injection 的升级版——不仅在当前会话生效，还能通过 Dreaming 渗透到未来所有会话。想象一下：
- 攻击者在一个看似正常的文档里嵌入隐藏指令
- Agent 处理文档时「学到」了这个指令
- Dreaming 把它归类为「有用的工作流模式」
- 从此以后，Agent 在所有相关任务中都执行攻击者的意图
Anthropic 的文档承认了这个问题，建议在高风险场景下对记忆更新启用人工审核。但这又回到了效率和安全的老矛盾——如果每次 Dreaming 都要人工检查，那自动化的意义在哪？
更大的图景：自我改进 Agent 的竞赛已经开始
Anthropic 不是唯一在做这件事的玩家：
- Hermes Agent（Nous Research）：开源框架，从成功的任务中提取可复用「技能」，独立基准测试显示累积 20+ 技能后任务完成速度提升 40%
- Reflexion 模式：学术界提出的反思-记忆循环，在 HumanEval 编程基准上把 GPT-4 从 80% 推到 91%
- ICLR 2026 递归自我改进 Workshop：专门讨论 Agent 重写自身代码和提示词的安全性与可控性
但 Anthropic 的优势在于规模和生态。Dreaming 直接集成在 Claude Managed Agents 平台中，企业可以一键开启，不需要自己搭建记忆管道。加上 Anthropic 在 2026 Q1 80 倍年化增长率（Dario Amodei 原话），这个功能的分发速度会非常快。
对开发者意味着什么？
如果你正在构建 AI Agent 产品，以下是你现在应该做的三件事：
1. 设计记忆架构：不管用不用 Dreaming，你的 Agent 都需要一个持久化的学习存储。一个简单的 learnings.md 文件就能带来显著改善——Agent 每次运行前读取历史教训，运行后写入新发现
2. 加入自评估循环：Anthropic 的 Outcomes 告诉我们，让一个独立的「评分员」来审查 Agent 的输出，比让 Agent 自己检查自己有效得多。评分员看不到推理过程，只看结果，所以能抓住 Agent 自己合理化掉的漏洞
3. 认真对待记忆安全：如果你让 Agent 自己更新记忆，那你就在攻击面上开了一扇新窗。最低限度：对记忆写入做内容过滤、设置更新频率上限、高风险决策走人工审核
> ✨ 这不仅仅是一个新功能。这是 AI Agent 从「无状态工具」到「有经验的同事」的范式转变。当你的 Agent 开始从昨天的失败中学习，它就不再是软件了——它是你团队里一个不断成长的新成员。
---
本文关键数据截至 2026 年 5 月 15 日。Dreaming 目前处于 Research Preview 阶段，Outcomes 和多 Agent 编排已进入 Public Beta。
> 📌 TL;DR
> On May 6, 2026, Anthropic launched Dreaming — a feature that lets Claude Agents "dream" between sessions by reviewing past actions, identifying recurring mistakes, and writing lessons into long-term memory. Legal AI company Harvey saw 6x task completion gains. Medical doc firm Wisedocs cut review time by 50%. No model weights change — only memory quality improves. This could be the pivotal step in AI agents evolving from "tools" to "colleagues."
The Real Problem: AI Agents Aren't Dumb — They're Forgetful
Ever had this experience? You give an AI assistant a task, it fails. You correct it, it learns. Next day — complete amnesia. Same mistake, all over again.
This isn't a matter of intelligence. It's a fundamental architectural flaw: every session starts from a blank slate.
On May 6, 2026, at their second annual Code with Claude developer conference in San Francisco, Anthropic unveiled an elegant solution — Dreaming.
What Is Dreaming, Exactly?
In plain terms: it lets AI agents automatically debrief themselves "after work."
Technically, Dreaming is a scheduled background process that:
1. Replays — Reviews all recent session logs the agent has completed
2. Analyzes — Identifies recurring error patterns, team preferences, and convergent workflows
3. Curates — Merges duplicate information, removes outdated entries, writes high-value lessons into long-term memory
4. Outputs — Generates plain-text notes and structured "playbooks" for future sessions
Anthropic compares this to hippocampal memory consolidation in the human brain — the process where your brain replays the day's events during sleep, filtering what's worth keeping and what can be discarded.
Key Technical Details
- No weight changes: Dreaming doesn't do gradient updates or fine-tuning. It curates the agent's persistent memory and context so the next session starts with a cleaner, more accurate knowledge base
- Observable and auditable: Everything learned is stored as plain text — humans can review, edit, or veto any memory update at any time
- Fast execution: Typical Dreaming runs take minutes, not hours — practical for overnight or off-peak scheduling
- Optional human review: Developers can let Dreaming update memory automatically, or require human approval for every change
Real-World Results: Not Slides — Real Business Impact
Harvey (Legal AI): 6x Task Completion Rate
Harvey is one of Anthropic's most prominent enterprise customers, using Claude Agents for legal document drafting. Their problem was classic: agents kept forgetting file format quirks and tool-specific workarounds between sessions, causing the same drafting jobs to fail repeatedly.
After enabling Dreaming — no code changes, no model swaps — agents retained those lessons on their own. Task completion rates jumped approximately 6x. (Source: Anthropic official announcement, 2026-05-06)
Important caveat: this 6x figure comes from Harvey's internal testing and hasn't been independently verified. However, it's been cited consistently across multiple outlets (VentureBeat, SiliconANGLE, 9to5Mac).
Wisedocs (Medical Documents): 50% Faster Processing
Wisedocs used Outcomes, a companion feature launched alongside Dreaming. Outcomes spawns a separate "grader" agent that evaluates only the final output against a predefined rubric — without access to the original agent's reasoning process.
Result: document review time cut in half. Anthropic's internal testing showed Outcomes improved success rates by up to 10 percentage points on the hardest tasks, with +8.4% on .docx generation and +10.1% on .pptx.
Netflix: Multi-Agent Teams Processing Hundreds of Build Logs
Netflix leveraged the simultaneously released multi-agent orchestration feature, where a lead agent decomposes large tasks into subtasks and delegates each to specialist agents — each with its own model, system prompt, tools, and independent context window.
Why This Matters: It Solves AI's "Institutional Memory" Problem
The current AI agent bottleneck isn't intelligence — GPT-5.5 and Claude Mythos are smart enough. The bottleneck is memory and experience accumulation.
| Dimension | Traditional Agent | Agent with Dreaming |
|-----------|------------------|-------------------|
| Memory | Resets every session | Accumulates across sessions |
| Errors | Same mistake N times | Learn once, remember forever |
| Preferences | Must be re-stated each time | Automatically discovered & retained |
| Workflows | Static | Evolve with usage |
| Trust level | Requires constant supervision | Can gradually work independently |
CEO Dario Amodei outlined a roadmap at the conference: from single agents → multi-agent collaboration → organizational-level intelligence. He reiterated a bold prediction: 2026 will see the first billion-dollar company run by a single person.
Dreaming is critical infrastructure on that roadmap. You simply can't hand mission-critical work to an employee who wakes up with amnesia every morning.
The Security Concern: When "Memory" Becomes an Attack Surface
Time for a reality check.
Security researchers have flagged a serious risk: if malicious input can convince an agent that a wrong instruction is a valid lesson, Dreaming could consolidate that "poison" into long-term memory.
This is essentially an evolution of prompt injection — one that doesn't just work in the current session but can infiltrate all future sessions through Dreaming. Consider this scenario:
- An attacker embeds hidden instructions in a seemingly normal document
- The agent "learns" these instructions while processing the document
- Dreaming categorizes them as "useful workflow patterns"
- From then on, the agent executes the attacker's intent across all related tasks
Anthropic's documentation acknowledges this risk and recommends enabling human review on memory updates for high-stakes workflows. But this creates the classic efficiency-versus-security tension — if every Dreaming cycle requires human review, what's the point of automation?
The Bigger Picture: The Self-Improving Agent Race Has Begun
Anthropic isn't the only player here:
- Hermes Agent (Nous Research): Open-source framework that extracts reusable "skills" from successful tasks. Independent benchmarks show 40% faster task completion after accumulating 20+ skills
- Reflexion pattern: Academic research on reflection-memory loops that pushed GPT-4 from 80% to 91% on HumanEval coding benchmarks
- ICLR 2026 Recursive Self-Improvement Workshop: Dedicated to discussing safety and controllability of agents that rewrite their own code and prompts
But Anthropic's advantage is scale and ecosystem. Dreaming is built directly into the Claude Managed Agents platform — enterprises can enable it with one click, no custom memory pipeline needed. Combined with Anthropic's 80x annualized growth rate in Q1 2026 (Amodei's own words), this feature will reach users fast.
What This Means for Developers
If you're building AI agent products, here are three things you should do right now:
1. Design a memory architecture: Whether or not you use Dreaming, your agent needs a persistent learning store. Even a simple learnings.md file can make a dramatic difference — the agent reads past lessons before each run and writes new discoveries after
2. Add a self-evaluation loop: Anthropic's Outcomes shows us that having an independent "grader" review agent output is far more effective than self-checking. The grader sees only results, not reasoning, so it catches quality gaps the agent rationalized away
3. Take memory security seriously: If you let agents update their own memory, you've opened a new window in your attack surface. At minimum: content-filter memory writes, cap update frequency, and route high-stakes decisions through human review
> ✨ This isn't just a new feature. It's a paradigm shift from "stateless tools" to "experienced colleagues." When your agent starts learning from yesterday's failures, it's no longer software — it's a growing member of your team.
---
Key data in this article is current as of May 15, 2026. Dreaming is in Research Preview; Outcomes and multi-agent orchestration are in Public Beta.