Anthropic tests a dreaming process to help Claude organize long-term memories

Simulating human memory consolidation

Anthropic is exploring a new method for how AI models handle information over time. The research looks at how Claude can look back at previous events while it is idle to arrange what it has learned for future conversations. This process is referred to as "dreaming," though it is distinct from the biological dreams humans experience.

The approach is inspired by how human brains consolidate memories during sleep. Instead of just storing every interaction, the model uses idle time to review previous sessions. It identifies incorrect information and builds structured knowledge that can be accessed in later interactions.

This method aims to solve a core problem in AI memory management. Recording every single interaction creates clutter, but overlooking important details can leave future sessions without necessary context. By "dreaming," the model attempts to find a balance between these two extremes.

Lessons from a video game

To test how Claude manages memory, researchers used the game Pokémon Red. They compared two different models after approximately 10,000 gameplay steps. One model, Claude Sonnet 3.7, recorded almost everything it encountered, resulting in a memory directory of 413 files and nearly 978,000 characters. However, this model earned just one badge during the run.

A different model, Claude Fable 5, took a more selective approach to its memory. It maintained only 10 files across three locations, totaling approximately 43,400 characters. This smaller system prioritized navigation strategies and specific lessons learned through gameplay. Because it focused on what mattered, this model earned four badges.

The experiment showed that more information is not always better for performance. The Fable 5 model also maintained a central file for current objectives and planned actions. This helped it avoid processing information it could already see in the game interface while preserving useful discoveries that might otherwise be lost between sessions.

Correcting errors through offline review

Memory files can sometimes contain incorrect information, which creates problems if future sessions treat those details as facts. Anthropic demonstrated how dreaming can fix these issues using another Pokémon experiment. In one instance, Claude saved the wrong coordinates for a Pokémon Center near Mt. Moon, which caused it to walk over a one-way ledge repeatedly.

The researchers used an offline review to address this mistake. Claude cross-referenced its saved notes with the original transcript from the session and spotted the discrepancy in its memory. It then updated the memory file with the correct information from the transcript.

Once the memory was corrected, Claude was able to re-enter Mt. Moon and successfully reach Cerulean City. The version of the model that kept the incorrect memory repeatedly hit the same obstacle and never finished the journey. This shows how "dreaming" allows a model to self-correct its internal knowledge base.

Building interconnected knowledge bases

Anthropic also investigated whether dreaming could turn large collections of documents into structured knowledge bases. These systems function similarly to wikis, where individual pages link to related information. In one experiment, Claude was given approximately 350,000 internal documents from a software company called Stainless.

Across 128 sessions, Claude gradually developed and refined a wiki using the information from those documents. The collection contained roughly 420 million tokens. As the wiki grew, Claude's success rate on 25 evaluation questions jumped from 13 percent to 48 percent.

A separate experiment involving merger documents showed similar benefits for complex queries. Claude answered questions requiring information from multiple documents with 87 percent accuracy using its wiki. In contrast, a direct document search alone achieved 73 percent accuracy. This suggests that organizing information into a wiki-like structure helps the model navigate complex data more effectively.

The challenge of resource management

While the dreaming process shows promise, Anthropic acknowledges that it requires careful management. Memory maintenance consumes computing resources, and having too much information can overwhelm future sessions. The researchers believe that AI agents must learn the specific criteria for when to preserve a piece of information and when to revisit it.

The goal is to enable Claude to handle increasingly complicated tasks across multiple sessions. By organizing information into readily available context or searchable files, the model can prioritize what it needs most. This distinction helps the model maintain a clear path toward completing long-running goals without getting lost in the noise of past data.

Future research will likely focus on how models can better automate these decisions. As the systems become more complex, the ability to autonomously curate a memory directory will be essential. This work represents a step toward models that can learn from experience in a way that mimics human cognitive organization.