Originally published on the Arize AI blog: Two labs started dreaming, and they built two different architectures.最初发表于Arize AI博客：两个实验室开始做梦，构建了两种不同的架构。
On May 6, 2026, Anthropic launched Dreams for its Managed Agents platform, the system for hosting stateful background agents. 27 days later, on June 2, OpenAI shipped Dreaming V3 for ChatGPT, the third iteration of a feature it has called “dreaming” since April 2025. By shipping under the same word, Anthropic adopted a technical term OpenAI had been using for 14 months. Both companies now treat the release as central to a multi-year arc Sam Altman has described publicly since 2025 as memory approaching the shape of remembering your whole life.2026年5月6日，Anthropic为其Managed Agents平台推出了Dreams，该平台用于托管有状态的后台代理。27天后，即6月2日，OpenAI为ChatGPT发布了Dreaming V3，这是自2025年4月以来被称为“做梦”功能的第三次迭代。通过使用同一个词发布，Anthropic采用了OpenAI已使用14个月的技术术语。两家公司现在都将此次发布视为多年发展弧线的核心，Sam Altman自2025年以来公开描述过这一弧线，即记忆逐渐接近记住你整个人生的形态。
Dreaming, in the agent-memory sense, is the pattern where a background process reads past session transcripts and writes a consolidated memory artifact the next session can use. It has a 40-year lineage in machine learning, but the productized form is new. Underneath this shared word, Anthropic and OpenAI built two different architectures. The research that landed in the same window suggests the difference matters more than the convergence does.在代理记忆的意义上，“做梦”是一种模式，其中后台进程读取过去的会话记录，并编写一个合并的记忆工件供下一个会话使用。它在机器学习中有40年的历史，但产品化形式是新的。在这个共享词汇下，Anthropic和OpenAI构建了两种不同的架构。在同一时期出现的研究表明，这种差异比趋同更重要。
This post is about what each lab actually shipped, what the empirical literature says about the consolidation pattern in general, and what to do if you are the team trying to figure out whether your version of dreaming is helping the agent or quietly degrading it.这篇文章将介绍每个实验室实际发布的内容、关于合并模式的实证文献，以及如果你的团队试图弄清楚你的“做梦”版本是在帮助代理还是在悄悄降低其性能时该怎么做。
What Anthropic shippedAnthropic发布的内容
Anthropic’s Dreams is the more cautious of the two designs, and its caution is the most interesting thing about it. Anthropic launched the feature as a research preview at its Code with Claude developers’ conference on May 6, scoped to Managed Agents.Anthropic的Dreams是两种设计中较为谨慎的一种，而其谨慎之处正是最有趣的地方。Anthropic于5月6日在Code with Claude开发者大会上以研究预览形式推出了该功能，范围限定于Managed Agents。
The API is explicit. You call it asynchronously, hand it an existing memory store and between one and 100 raw session transcripts, choose a model, and gate the whole thing behind a beta API header. The dream produces a separate, reorganized memory store, with the input store left alone. The Anthropic documentation is unusually direct about this commitment:API是明确的。你异步调用它，传入一个现有的记忆存储和1到100个原始会话记录，选择一个模型，并将整个过程放在一个beta API头后面。这个“梦”会产生一个独立的、重新组织的记忆存储，而输入存储保持不变。Anthropic的文档对这一承诺异常直接：
“A dream reads an existing memory store alongside past session transcripts, then produces a new, reorganized memory store… The input store is never modified.”“一个‘梦’会读取现有的记忆存储以及过去的会话记录，然后生成一个新的、重新组织的记忆存储……输入存储永远不会被修改。”
The commitment is doing architectural work: Anthropic’s engineers expect the dream to sometimes degrade the memory rather than improve it, and they have built the system so that the comparison is possible. You can run an eval against the pre-dream store, run the same eval against the post-dream store, and decide whether to promote the new artifact or throw it away. The dream is an experiment, not a commitment.这一承诺具有架构意义：Anthropic的工程师预计“梦”有时会降低记忆质量而非改善，因此他们构建的系统使得比较成为可能。你可以对“梦”前的存储运行评估，再对“梦”后的存储运行相同评估，然后决定是推广新工件还是丢弃它。“梦”是一个实验，而非承诺。
The commit boundary is also a choice the team running Dreams gets to make. From Ars Technica’s reporting on the launch: “Users will be able to choose between an automatic process, or reviewing changes to memory directly.” You can let the system auto-promote the new store, or you can require human review before changes take effect. Either way the previous state is preserved.提交边界也是运行Dreams的团队可以做出的选择。据Ars Technica对发布的报道：“用户将能够在自动流程或直接审查记忆更改之间进行选择。”你可以让系统自动推广新存储，也可以要求人工审核后再生效。无论哪种方式，之前的状态都会被保留。
Anthropic’s own framing of the value is cross-agent rather than per-user. Dreaming “restructures memory so it stays high-signal” across team-shared workflows. The interesting word is “restructures.” Anthropic believes the dream is constructive; they have just chosen to leave the previous structure intact so a team can fall back to it.Anthropic自己对价值的定位是跨代理而非每个用户。做梦“重构记忆以保持高信号”，适用于团队共享的工作流程。有趣的是“重构”这个词。Anthropic相信“梦”是建设性的；他们只是选择保留之前的结构，以便团队可以回退。
The system reads as if it were designed by people who had thought hard about what could go wrong.这个系统看起来像是那些深思熟虑过可能出错情况的人设计的。
What OpenAI shippedOpenAI发布的内容
OpenAI’s Dreaming for ChatGPT is the more ambitious of the two designs. Dreaming is not a new feature: “Dreaming V0” was introduced in April 2025 alongside saved memories, and what shipped on June 2 is Dreaming V3, described in the launch post as “a significantly more capable and compute-efficient memory architecture built on top of dreaming.”OpenAI的ChatGPT Dreaming是两种设计中更为雄心勃勃的一种。Dreaming并非新功能：“Dreaming V0”于2025年4月与保存的记忆一同推出，而6月2日发布的是Dreaming V3，在发布文章中描述为“基于做梦构建的、能力显著更强且计算效率更高的记忆架构”。
OpenAI describes the mechanism in its own words: “dreaming leverages a background process that allows ChatGPT to learn from many conversations and synthesize ChatGPT’s memory state in order to always provide the freshest, most relevant context to your conversations.” The companion memory FAQ uses parallel language: “ChatGPT’s memory is based on a continually updated synthesis of context from your past chats.”OpenAI用自己的话描述了这一机制：“做梦利用后台进程，让ChatGPT从许多对话中学习，并综合ChatGPT的记忆状态，以便始终为你的对话提供最新、最相关的上下文。”配套的记忆FAQ使用了类似的语言：“ChatGPT的记忆基于对你过去聊天内容的持续更新综合。”
Both documents describe a single canonical state that gets rewritten as the system runs, not a parallel artifact alongside the existing one. “Continually updated synthesis” implies that each pass is built on whatever the previous one wrote, though OpenAI does not state this explicitly.两份文档都描述了一个单一的规范状态，该状态随着系统运行而被重写，而不是与现有状态并行的工件。“持续更新综合”意味着每次传递都基于前一次写入的内容，尽管OpenAI并未明确说明这一点。
OpenAI presents Dreaming as solving temporal supersession. “You’re going to Singapore in July” should become “You went to Singapore in July 2026” once the trip ends, automatically, without manual intervention.OpenAI将Dreaming描述为解决时间更替问题。“你七月要去新加坡”应该在旅行结束后自动变为“你2026年七月去了新加坡”，无需人工干预。
The user-facing controls are the closest thing the architecture has to an escape hatch. A memory summary updates roughly hourly (the FAQ documents an “Updated 2h ago” indicator), and users can edit individual entries or hide them. There’s also a settings toggle to revert to legacy saved memories.用户控制是架构中最接近逃生舱的东西。记忆摘要大约每小时更新一次（FAQ中显示“2小时前更新”的指示器），用户可以编辑单个条目或隐藏它们。还有一个设置开关可以恢复到传统的保存记忆。
Those controls operate on individual items, not on the synthesized state itself. The memory FAQ is explicit about the distinction: “details from past chats can change over time as ChatGPT updates what’s more helpful to remember. Because ChatGPT doesn’t retain every detail from past chats, use saved memories for anything you want it to always remember.” Saved memories (the explicit 2024 layer, with per-entry edit and restore-by-date) are the persistent surface. The Dreaming synthesis is the mutable one. The legacy-memory toggle switches between the two systems; it does not roll back a dream.
这些控制作用于单个项目，而非综合状态本身。记忆FAQ明确区分了这一点：“过去聊天中的细节可能会随着ChatGPT更新更值得记住的内容而改变。由于ChatGPT不会保留过去聊天中的每个细节，请使用保存的记忆来存储你希望它始终记住的内容。”保存的记忆（明确的2024层，支持逐条编辑和按日期恢复）是持久表面。Dreaming综合是可变的。传统记忆开关在两个系统之间切换；它不会回滚一个“梦”。
The summary itself is a partial view: the FAQ adds that it “will not include everything that ChatGPT remembers based on your chats.”摘要本身是部分视图：FAQ补充说它“不会包含ChatGPT根据你的聊天记住的所有内容。”
The system reads as if it were designed by people who believed the dreaming metaphor.这个系统看起来像是那些相信“做梦”隐喻的人设计的。
The divergence分歧
Anthropic preserves the input store and produces a parallel artifact. OpenAI rewrites the memory in place and lets the next pass condition on the previous one’s output. Both companies use the same word for the feature, but they made opposite bets about whether the dream should be reversible.
Anthropic保留输入存储并生成并行工件。OpenAI就地重写记忆，并让下一次传递基于前一次的输出。两家公司对该功能使用了相同的词，但他们对“梦”是否可逆做出了相反的赌注。
This is not a small distinction. The empirical literature on iterated LLM-driven consolidation, which landed in the same window as both launches, suggests it is the most important distinction in the entire feature.这不是一个小区别。关于迭代LLM驱动合并的实证文献，与两次发布同期出现，表明这是整个功能中最重要的区别。
Why the divergence matters: Zhang’s paper分歧为何重要：张的论文
In May 2026, Dylan Zhang at the University of Illinois Urbana-Champaign published Useful Memories Become Faulty When Continuously Updated by LLMs (project page, arXiv 2605.12978). The paper is the cleanest empirical case yet made against the iterated-consolidation pattern, and the central finding is striking.
2026年5月，伊利诺伊大学厄巴纳-香槟分校的Dylan Zhang发表了《有用的记忆在持续更新时变得有缺陷》（项目页面，arXiv 2605.12978）。这篇论文是对迭代合并模式最清晰的实证反驳，其核心发现令人震惊。
On ARC-AGI, the abstraction-and-reasoning benchmark, GPT-5.4 solves 19 problems at 100% accuracy without memory. After streaming those same problems through consolidation loops with ground-truth solutions available, accuracy drops to 54%.在抽象与推理基准ARC-AGI上，GPT-5.4在没有记忆的情况下以100%的准确率解决了19个问题。在通过合并循环流式传输相同问题（并提供真实答案）后，准确率降至54%。
The trajectories were perfect. The failure happened in the rewrite step. The act of compressing right answers into a re-usable lesson, Zhang argues, is what made the system forget how to solve them.轨迹是完美的。失败发生在重写步骤。张认为，将正确答案压缩成可重复使用的教训这一行为，正是导致系统忘记如何解决它们的原因。
The pattern repeats across benchmarks.这种模式在多个基准上重复出现。
- On ScienceWorld, a text-based scientific reasoning benchmark, score peaks after roughly 20 updates and then declines below the no-memory baseline by step 100.在基于文本的科学推理基准ScienceWorld上，分数在大约20次更新后达到峰值，然后在第100步时下降到低于无记忆基线。
- On WebShop, a simulated e-commerce task suite, performance falls from 0.64 with eight examples to 0.20 with 128, matching the no-memory baseline at scale.在模拟电子商务任务套件WebShop上，性能从8个示例时的0.64下降到128个示例时的0.20，在大规模下与无记忆基线持平。
- On ALFWorld, an embodied household-task benchmark, three different solver sizes (Qwen 27B, 9B, and 4B) all show utility decay across consolidation steps.在具身家务任务基准ALFWorld上，三种不同规模的求解器（Qwen 27B、9B和4B）在合并步骤中都显示出效用衰减。
Zhang identifies three reproducible failure modes inside the consolidation step:张在合并步骤中识别出三种可复现的失败模式：
- Misgrouping – episodes from distinct classes merge into single entries错误分组——来自不同类别的片段合并为单个条目
- Interference – abstraction strips applicability conditions, so lessons generalize too broadly干扰——抽象剥离了适用条件，导致教训泛化过广
- Overfit – specific selectors get deleted as the input distribution narrows过拟合——随着输入分布变窄，特定选择器被删除
The architectural diagnosis Zhang offers is precise:张提供的架构诊断很精确：
“Continuously updated textual memory is an iterated generative loop with no anchor… it is a sample — fluent, confident, and increasingly disconnected from what actually happened.”“持续更新的文本记忆是一个没有锚点的迭代生成循环……它是一个样本——流畅、自信，且越来越脱离实际发生的情况。”
Two memory modes appear in Zhang’s experiments:张的实验中出现了两种记忆模式：
- Episodic memory retains the raw rollouts: the actual session transcripts, unchanged, and the system retrieves them at recall time.情景记忆保留原始记录：实际的会话记录，保持不变，系统在回忆时检索它们。
- Abstract memory stores the LLM’s compressed summary of those rollouts, the lessons and generalized rules synthesized from them, and discards the originals.抽象记忆存储LLM对这些记录的压缩摘要，从中综合出的教训和泛化规则，并丢弃原始记录。
Most production memory systems, including both Dreams and Dreaming, sit on the abstract end.大多数生产记忆系统，包括Dreams和Dreaming，都处于抽象端。
Zhang’s recommendation is architectural: episodic-only memory recovers nearly the entire gain consolidation-based variants give up. Abstract-only never beats the no-memory baseline. The implication is sharp: the abstraction step is the failure point, and both Dreams and Dreaming run it. Only OpenAI runs it iteratively against its own previous output, which is the specific pattern Zhang’s results target.张的建议是架构性的：仅情景记忆几乎恢复了基于合并的变体所放弃的全部增益。仅抽象记忆从未超过无记忆基线。其含义很尖锐：抽象步骤是失败点，而Dreams和Dreaming都运行它。只有OpenAI迭代地针对自身之前的输出运行它，这正是张的结果所针对的具体模式。
The paper is not a rejection of memory as a capability. It is a rejection of the specific pattern of letting an LLM rewrite its own memory in place across many iterations. That pattern is exactly what OpenAI’s Dreaming architecture does, while Anthropic’s Dreams architecture does not.这篇论文并非拒绝记忆作为一种能力。它拒绝的是让LLM在多次迭代中就地重写自己记忆的具体模式。这种模式正是OpenAI的Dreaming架构所做的，而Anthropic的Dreams架构则没有。
How the architectures meet the research架构如何应对研究
Anthropic’s design preserves the input store. That means a team running Anthropic Dreams can run an eval against the pre-dream store, run the same eval against the post-dream store, and discard the post-dream artifact if it scores worse. The dream is reversible. The iterated-generative-loop failure mode Zhang documents is constrained to whatever happens inside a single dream invocation; across invocations, the team chooses whether to keep what the dream produced or roll back to the previous state.Anthropic的设计保留了输入存储。这意味着运行Anthropic Dreams的团队可以对“梦”前的存储运行评估，再对“梦”后的存储运行相同评估，如果“梦”后的工件得分更差则丢弃它。“梦”是可逆的。张记录的迭代生成循环失败模式被限制在单次“梦”调用内部；在多次调用之间，团队可以选择保留“梦”产生的内容还是回滚到之前的状态。
OpenAI’s design rewrites in place. The previous state is overwritten, and the per-item user controls (edit, “don’t mention this again”) are not a pre/post comparison surface. That shape is the one Zhang identifies as failing.OpenAI的设计就地重写。之前的状态被覆盖，而逐条用户控制（编辑、“不再提及”）不是前后比较的表面。这种形态正是张识别为失败的那种。
TechTimes’ coverage of the launch put the same point in user-facing terms: the redesign “limits the granular audit trail that the discrete saved-memories list provided.”TechTimes对发布的报道用面向用户的术语表达了同样的观点：重新设计“限制了离散保存记忆列表提供的细粒度审计追踪。”
OpenAI’s framing creates a particular tension with Zhang’s findings. OpenAI claims Dreaming handles temporal supersession. Old facts get gracefully replaced as time passes, the Singapore trip going from upcoming to over to historical without manual intervention. Zhang’s mechanism suggests iterated consolidation is particularly bad at preserving the specific facts that need to be updated. The Singapore trip is exactly the kind of detail (a specific date, a specific place) that Zhang’s failure mode predicts will get rounded off across passes into something vaguer and less actionable. Whether OpenAI’s Dreaming actually preserves and updates temporal specifics is empirically open. The architecture doesn’t structurally guarantee it.OpenAI的框架与张的发现产生了特定的紧张关系。OpenAI声称Dreaming处理时间更替。旧事实随着时间推移被优雅地替换，新加坡之旅从即将到来变为结束再到历史，无需人工干预。张的机制表明，迭代合并特别不擅长保留需要更新的具体事实。新加坡之旅正是那种细节（具体日期、具体地点），张的失败模式预测它会在多次传递中被舍入为更模糊、更不可操作的内容。OpenAI的Dreaming是否真正保留和更新时间细节在经验上是开放的。架构在结构上并不保证这一点。
We do not yet have enough data to know whether OpenAI’s implementation produces the kind of degradation Zhang documents. What we know is that the architectural shape of the system resembles the shape Zhang’s paper says fails, and the design does not include the escape hatch Anthropic’s design does.我们还没有足够的数据来知道OpenAI的实现是否会产生张记录的那种退化。我们知道的是，系统的架构形态类似于张论文中声称会失败的形态，并且设计中没有包含Anthropic设计中的逃生舱。
That is not an accusation of bad engineering. It is an observation about which architecture left itself the option to find out.这不是对糟糕工程的指责。这是一个关于哪种架构给自己留下了发现问题的选项的观察。
What to do if you are shipping dreaming如果你正在发布“做梦”功能该怎么做
If you are a team building memory consolidation into an agent, the practical takeaway from the launches and the research is straightforward.
如果你是一个正在为代理构建记忆合并的团队，从发布和研究中得出的实际结论很直接。
Preserve the input – Whatever your consolidation step does, write the result to a new artifact and keep the previous state. The cost at the storage layer is small, and the upside is the ability to compare and roll back.保留输入——无论你的合并步骤做什么，将结果写入新工件并保留之前的状态。存储层的成本很小，而好处是能够比较和回滚。
Instrument the loop – Without traces of every consolidation invocation, every retained or discarded artifact, and every session that runs after, you have nothing to evaluate against. This is the work we have been doing on our own agent, Alyx. Consolidation-shaped operations leave traces in Arize, and golden sessions captured from production traces become the eval set.对循环进行检测——如果没有每次合并调用、每个保留或丢弃的工件以及每个后续会话的追踪，你就没有可评估的内容。这是我们自己在代理Alyx上所做的工作。合并形状的操作在Arize中留下追踪，而从生产追踪中捕获的黄金会话成为评估集。
Compare with an eval – Define a measurable behavior the agent is supposed to do in the next session: answer a question, recall a preference, complete a task. Run it against the pre-dream store and the post-dream store. Without the comparison, you are trusting the LLM’s instinct for what a good lesson looks like, which Zhang’s paper says is the wrong thing to trust.通过评估进行比较——定义代理在下一个会话中应该执行的 measurable 行为：回答问题、回忆偏好、完成任务。对“梦”前存储和“梦”后存储运行评估。没有比较，你就是在信任LLM对什么是一个好教训的本能，而张的论文说这是不该信任的。
Plan to forget – Consolidation is additive by default. Decide what your system drops, not just what it keeps. A memory store that only grows is a memory store that eventually fills the context window with noise. Selective forgetting is the operation that protects the rest.计划遗忘——合并默认是累加的。决定你的系统丢弃什么，而不仅仅是保留什么。一个只增长的记忆存储最终会用噪声填满上下文窗口。选择性遗忘是保护其余部分的操作。
Zhang’s paper gave us the empirical test: an LLM that rewrites its own memory across iterations forgets how to do things it had once solved. Memory engineering in 2026 means taking that test seriously: preserve the input, instrument the loop, compare with an eval, plan to forget, and treat consolidation as an experiment rather than a commitment. Anthropic’s architecture lets you do this. OpenAI’s architecture does not, which means the teams who care about whether their dreaming is working will have to build the comparison themselves before they trust the product to do anything load-bearing.张的论文给了我们实证测试：一个在多次迭代中重写自己记忆的LLM会忘记它曾经解决过的事情。2026年的记忆工程意味着认真对待这个测试：保留输入、检测循环、通过评估进行比较、计划遗忘，并将合并视为实验而非承诺。Anthropic的架构允许你这样做。OpenAI的架构不允许，这意味着关心自己“做梦”是否有效的团队必须在信任产品承担任何重要任务之前，自己构建比较机制。
Two architectures of dreaming now ship at scale. The metaphor converged, the implementation did not. This divergence is the thing to watch.两种“做梦”架构现在大规模发布。隐喻趋同了，实现却没有。这种分歧是值得关注的事情。