Can agents dream? Anthropic says yes
In an impressive sequence of releases, another new feature has arrived in Claude: Dreams. According to Anthropic, this gives the agent the ability to “dream.”
Memory Stores
Throughout different conversational sessions, agents store valuable information that can be reused later. Called Memory Stores, these items include summaries, preferences, project conventions, errors and issues, domain context, and more.
With these stores, the knowledge from one session is carried into another in an optimized, concise way so that the new context window knows only what is necessary.
The new feature aims to solve a common problem: invalid or outdated information. Whether due to changed domain rules, old data, or contradictory entries, memory can accumulate inaccuracies.
Claude Dreams
Anthropic’s agents now, asynchronously, can capture up to 100 previous sessions and gather transcripts, logs and relevant information to clean their Memory Stores, producing a new “report” as the result of that cleanup.
In other words, the agent can now assess which information is contradictory, what should be removed due to duplication, and which new insights emerge from that analysis.
The most important point is that the original storage is not altered directly. The system generates a separate output that can be reviewed, accepted or discarded by the developer.
Access and usage
This tool is in Research Preview, and you can request access here: Claude Dreams.
Dreams are billed as normal API usage and can be used with models like Opus 4.7 and Sonnet 4.6.
