Anthropic has revealed how its Claude AI assistant uses a process called “Dreaming” to organise memory, correct previous mistakes and improve its performance across multiple sessions. Inspired by the way human brains consolidate memories during sleep, the system allows AI agents to review past interactions during periods of inactivity, identify useful information and update their long-term memory.
The approach addresses one of the biggest challenges facing AI agents: retaining useful knowledge without accumulating outdated information, repeated notes and incorrect assumptions. Instead of treating every previous interaction as equally important, Claude can use Dreaming to identify patterns, verify earlier conclusions, create structured knowledge repositories and prepare relevant information for future tasks. Anthropic’s experiments suggest that this approach can improve performance on long-running activities that require an AI agent to remember what it has learned.
Key Takeaways
- Anthropic has explained how Claude uses Dreaming to improve long-term agent memory.
- Dreaming allows AI agents to review previous sessions and reorganise stored information between interactions.
- The process can correct inaccurate memories, remove duplicates and identify outdated information.
- Claude can build structured knowledge repositories, including interconnected wiki pages.
- In one database benchmark experiment, memory improved the performance of newer Claude models.
- In a Pokémon Red experiment, Dreaming corrected an incorrect navigation memory that repeatedly caused the AI agent to get stuck.
- In an experiment involving approximately 350,000 internal company documents, Dreaming helped improve knowledge retrieval.
- The process does not retrain Claude’s underlying model weights; it improves the external memory available to future sessions.
What Is Claude’s Dreaming Process?
Dreaming is an AI memory-management process developed by Anthropic for its Claude agents.
The idea is based on a distinction between collecting information and organising it.
During a normal task, an AI agent may encounter important information, make decisions, solve problems and occasionally make mistakes. It can store some of these experiences in memory files or session logs.
However, simply saving information does not guarantee that it will remain useful.
Over time, an agent’s memory can become cluttered with duplicate information, outdated instructions and incorrect assumptions. Important lessons may also become difficult to find among thousands of stored observations.
Dreaming attempts to solve this problem by allowing Claude to review its previous sessions outside the main task.
During this process, the AI can:
- Identify patterns across multiple conversations.
- Find mistakes in previously stored information.
- Compare memory notes with original session transcripts.
- Merge duplicate observations.
- Replace outdated information with more accurate details.
- Create new notes and structured knowledge repositories.
- Decide which information should be made available in future sessions.
The objective is to create a memory system that becomes more useful as an AI agent gains experience.
Why Does Anthropic Compare Dreaming to Human Sleep?
Human memory does not work like a simple recording device.
People encounter enormous amounts of information throughout the day, but only some experiences become long-term memories.
Researchers believe that processes occurring during sleep and quiet rest help consolidate information, strengthen important memories and integrate new experiences with existing knowledge.
Anthropic uses this as an inspiration for AI agent memory.
An AI agent can capture information while completing a task and then use a separate background process to review that information later.
The comparison is conceptual rather than biological. Claude does not sleep or experience human memory consolidation. Dreaming is a computational process that uses additional model inference to reorganise stored information.
How Claude Organises Its Memory
Anthropic describes Claude’s memory system as a hierarchy with four main layers.
Each layer serves a different purpose, allowing the agent to retain detailed experiences while keeping important information accessible.
| Memory layer | Purpose |
|---|---|
| Session transcripts | Preserve searchable records of previous interactions and tasks |
| Memory notes | Store important observations, verified facts and lessons |
| Memory representations | Organise knowledge into structured collections such as interconnected wikis |
| Injected context | Make selected information available automatically at the beginning of a new session |
The lowest layer contains the original interaction records.
Memory notes extract useful information from those records, while higher-level representations organise related knowledge into a structure that can be searched and reused.
The final layer contains selected information that can be loaded directly into a future session.
Dreaming helps maintain connections between these layers.
For example, if Claude discovers that a previously stored instruction is incorrect, it can return to the original transcript, verify what happened and update the corresponding memory note.
This reduces the risk of an earlier mistake influencing future decisions.
Dreaming Helps Claude Learn From Its Previous Mistakes
One of Anthropic’s most important experiments examined whether Claude could improve its performance by remembering incorrect answers and learning from the feedback.
The researchers used Continual Learning Bench, or CL-Bench, which evaluates how well an AI agent can use information from previous sessions to complete longer-running tasks.
In the experiment, Claude answered database questions.
After each answer, it received information about whether its response was correct. It could then save information in a persistent memory directory before attempting another question in a fresh session.
This created an opportunity for the AI to develop a learning process.
When Claude answered incorrectly, it could record the failure, investigate possible reasons and create a rule to avoid repeating the mistake.
Newer Claude Models Benefit From Persistent Memory
Anthropic found that different Claude models developed different approaches to maintaining memory.
Some models recorded unsuccessful attempts but did not consistently convert those experiences into verified rules.
More capable models began organising information into reference documents, tracking confirmed facts and identifying common mistakes.
Claude Opus 5 was among the first tested models to establish a more complete learning loop.
It recorded previous answers, documented verified rules and included warnings about mistakes that could affect future responses.
The performance comparison showed improvements when models could use persistent memory.
| Claude model | Without memory | With memory |
|---|---|---|
| Opus 5.5 | 0.655 | 0.838 |
| Sonnet 5.5 | 0.644 | 0.833 |
| Opus 5 | 0.564 | 0.828 |
| Opus 4.7 | 0.383 | 0.700 |
| Sonnet 5 | 0.306 | 0.610 |
| Sonnet 4.6 | 0.330 | 0.364 |
The results show that persistent memory improved performance across the tested models, although the size of the improvement varied.
The figures represent mean reward on the specific CL-Bench database task, not general-purpose accuracy across all AI activities.
The experiment demonstrates why memory can be particularly valuable for tasks where an agent encounters repeated problems and receives feedback over time.
A Pokémon Experiment Shows How Dreaming Can Correct Incorrect Memories
Anthropic also used Pokémon Red to demonstrate how incorrect memories can negatively affect an AI agent.
Although a video game may appear unrelated to professional AI applications, it provides a useful environment for testing long-running tasks.
The agent must remember its location, understand the map, plan routes and avoid repeating mistakes.
In one experiment, Claude was navigating Mt. Moon, a cave system in the game.
The AI incorrectly recorded that the Pokémon Center near the cave entrance was located on the eastern side of the mountain.
That incorrect information affected its navigation decisions.
After leaving the building, Claude believed it had successfully passed through the cave. It updated its memory to indicate that Mt. Moon was completed.
However, the agent had not actually reached the other side.
It subsequently walked over a one-way ledge and returned to the route from which it had entered.
The navigation tool did not clearly reveal what had happened, so Claude initially interpreted the problem incorrectly and continued relying on its inaccurate memory.
Dreaming Corrected the Navigation Error
Researchers then ran Dreaming on the agent’s stored memory and previous session transcript.
Claude compared its written notes with the original interaction record and identified the incorrect location information.
It subsequently corrected the memory.
With the updated information, the agent recognised that it needed to return to the cave entrance rather than continue towards the next town.
The corrected agent eventually reached Cerulean City.
The comparison was significant:
- With uncorrected memory, Claude repeatedly fell off the same ledge and returned to an earlier route.
- With Dreaming, it corrected the location information and successfully navigated through the cave.
- The run without memory also failed to complete the route.
This experiment demonstrates that an AI agent can become less effective when it stores incorrect information without a mechanism to verify it.
Dreaming provides a way to revisit the original evidence and repair that knowledge.
Claude Can Build Its Own Knowledge Wikis
Another important capability is the ability to organise information into interconnected knowledge repositories.
Anthropic calls these memory representations.
A wiki is one example. It consists of multiple pages containing related information, with links connecting topics and helping the agent navigate its knowledge.
This approach can be useful when an AI agent works with a large collection of documents over an extended period.
Instead of searching through every document from scratch during each session, the agent can build a structured map of important information.
Experiment With 350,000 Internal Documents
Anthropic tested this approach using approximately 350,000 internal documents from Stainless, a software company it acquired in May 2026.
The collection contained roughly 420 million tokens.
Researchers created an evaluation set containing 25 questions about the documents.
Claude agents were then instructed to explore the collection, build wiki pages, identify duplicate information and verify claims against original sources.
The experiment used multiple Dreaming sessions, with later sessions building on the work completed by earlier ones.
As the process expanded from two sessions to 128 sessions, performance on the 25-question evaluation increased from 13% to 48% of graded criteria when relying on the wiki alone.
The wiki also became more detailed.
| Metric | After 2 sessions | After 128 sessions |
|---|---|---|
| Wiki pages | 103 | 222 |
| Wiki links | 766 | 12,893 |
| Graded criteria met using wiki alone | 13% | 48% |
The results suggest that repeated memory organisation can help an AI agent develop a more useful understanding of a large document collection.
Combining Search With a Knowledge Wiki
Anthropic also examined whether a wiki could improve the performance of a conventional search tool.
The results showed that combining search with a growing wiki produced better performance than using search alone.
| Evaluation setup | Graded criteria met |
|---|---|
| Search tool without wiki | 53% |
| Wiki plus search after early development | 54% |
| Wiki plus search after extensive Dreaming | 61% |
| Wiki alone after 128 sessions | 48% |
The best result came from combining the wiki with search.
This is important because a wiki can provide an organised overview of information, while search allows the agent to retrieve specific source documents when it needs additional evidence.
Anthropic also reported that using both systems reduced token consumption by approximately 20,000 tokens per question compared with relying on search alone in the evaluated setup.
The findings suggest that structured memory can improve both information retrieval and the efficiency of long-running AI workflows.
Dreaming Can Decide What Information Claude Should Remember
Another challenge in AI memory design is deciding which information deserves to be stored.
An agent that saves every observation may eventually create an enormous collection of files that becomes difficult to manage.
However, an agent that saves too little information may repeatedly make the same mistakes.
Anthropic examined this problem through another Pokémon experiment.
Researchers compared memory created by Claude Sonnet 3.7 with memory produced by Claude Fable 5.
Sonnet 3.7 recorded a large number of individual events, including battles, items and encounters.
Its memory collection eventually contained 413 files across 225 folders.
Fable 5 took a different approach.
It created only nine memory files across two folders, along with a main file called global.md.
That file contained information about the current objective, plans for completing challenges and rules for using memory.
The agent prioritised strategies and information that would be useful in future situations instead of recording every event.
Fable 5 also progressed further in the game, earning four badges compared with one badge earned by Sonnet 3.7 at the point of comparison.
The experiment illustrates an important principle: effective AI memory is not necessarily about storing more information. It is about identifying what will be useful later.
How Dreaming Keeps Future Sessions More Efficient
Anthropic also described how selected memories can be made available automatically when a new session begins.
This process is called context injection.
Instead of requiring Claude to search for important information every time it starts a task, the agent can receive a small collection of relevant notes at the beginning of the session.
However, loading too much information can create another problem.
AI models have limited context windows, and unnecessary instructions can consume valuable space.
Anthropic therefore emphasises keeping the automatically loaded memory layer relatively small.
Additional information can remain in lower layers and be retrieved only when necessary.
Dreaming helps decide which information should be promoted into the immediately available context and which details should remain in searchable memory.
This approach aims to balance accessibility with efficiency.
What This Means for the Future of AI Agents
The research has implications for the development of AI agents that perform work over several hours, days or even months.
Traditional chatbots often treat each conversation as a relatively separate interaction.
More advanced AI agents need to retain knowledge about users, ongoing projects, previous decisions and the results of earlier tasks.
For example, a coding agent could remember which debugging approaches failed in a particular project.
A business research agent could maintain an organised knowledge base containing verified company information.
A personal assistant could retain useful preferences and learn from recurring corrections.
A research agent could build a structured record of previous experiments and their outcomes.
Dreaming offers one possible way to maintain these memories without requiring the agent to review its entire history during every interaction.
However, Anthropic acknowledges that memory maintenance introduces its own challenges.
Incorrect memories can still be created, information can become outdated and the Dreaming process consumes additional computing resources.
The company is continuing to investigate how frequently agents should review their memories and which information deserves additional verification.
The Bigger Picture
Anthropic’s Dreaming research highlights a major shift in AI development: future agents will need to become better at learning from their own work, not simply generating answers from existing knowledge. Persistent memory can help an agent retain useful experience, but without organisation and verification, it can also accumulate mistakes and outdated information. Dreaming attempts to address that problem by reviewing past sessions, correcting inaccurate notes and building structured knowledge that can be reused.
The longer-term opportunity is to create AI agents that become more effective as they work on recurring projects. Whether they are managing software development, analysing business documents or supporting scientific research, the ability to remember verified lessons could make them more useful over time. However, reliable memory will require careful control over accuracy, privacy, information freshness and computing costs. The next generation of AI agents may be defined not only by how well they reason in a single session, but by how effectively they learn from everything that came before.
Frequently Asked Questions
1. What is Claude’s Dreaming process?
Dreaming is an Anthropic-developed process that allows Claude agents to review previous sessions, reorganise stored memories, correct errors and create useful knowledge representations between interactions.
2. Why is it called Dreaming?
The name is inspired by theories of human memory consolidation during sleep. The AI process is computational and does not involve biological sleep.
3. Can Claude correct its own mistakes using Dreaming?
Yes. Claude can compare stored memory notes with original session transcripts, identify incorrect information and update its memory.
4. What is the difference between memory and Dreaming?
Memory stores information collected during an agent’s work. Dreaming reviews that information later, removes duplication, corrects errors and organises knowledge for future use.
5. What is Claude’s memory wiki?
It is a collection of interconnected knowledge pages that Claude creates and maintains. The wiki helps the agent organise information and retrieve relevant details across multiple sessions.
6. Did Dreaming improve Claude’s performance?
Yes. Anthropic reported improvements in its CL-Bench database experiment and in a document-retrieval experiment involving approximately 350,000 internal documents.
7. Does Dreaming retrain Claude’s AI model?
No. Dreaming organises external memory and creates information that future sessions can access. It does not directly update the underlying model weights.
8. Is Dreaming available to developers?
Anthropic introduced Dreaming for Claude Managed Agents as a research preview in May 2026. Its documentation describes an API-based process for reviewing memory stores and previous session transcripts.
Looking Ahead
Anthropic’s Dreaming research suggests that persistent memory could become one of the most important capabilities of advanced AI agents. As agents handle increasingly complex tasks, remembering verified information, identifying recurring mistakes and maintaining structured knowledge could help them work more effectively across sessions. The experiments show that these improvements can be measured in practical environments rather than only through conventional question-answering tests.
The challenge will be making this process reliable, efficient and secure. AI agents must know what to remember, when to update it and how to verify that stored information remains accurate. If companies can solve these problems, future AI assistants could become more capable of managing long-term projects and learning from previous work without repeatedly starting from scratch.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
