Anthropic introduced a new feature called “Dreams” for its Claude AI chatbot in May 2026, enabling AI agents to review past interactions and reorganize their memory stores over time. The feature was unveiled at the company’s Code with Claude developers’ conference.
The Dreams feature works by running an asynchronous background process in which Claude reads an existing memory store alongside transcripts from up to 100 past sessions. It then mines that data for patterns and insights, producing a reorganized memory store as output. Anthropic notes the process can take tens of minutes depending on the volume of data involved. Importantly, the original memory store is never modified — Claude generates a separate output that developers can review, keep, or discard.
Anthropic says the feature addresses a practical problem: as AI agents work with users across multiple sessions, their memory stores accumulate duplicate or redundant information. Dreams is designed to surface recurring patterns, common workflows, and shared preferences that a single agent session cannot identify on its own.
Developers retain control over how the feature operates. Dreams can update memory automatically or hold changes for manual review. Custom instructions can also be applied — for example, directing Claude to focus only on coding preferences while ignoring temporary debugging notes.
Legal AI company Harvey is cited as an early user, reporting a sixfold increase in task completion rates after using Dreams to help agents identify file-type workarounds and tool-specific patterns across drafting sessions.
Dreams currently supports Claude Opus 4.7 and Claude Sonnet 4.6 models and is available through beta headers on Anthropic’s Managed Agents platform. It is billed at standard API token rates, with costs scaling based on the number and size of sessions analyzed.
Anthropic described Dreams as part of a broader effort to build self-improving AI agents. The company also announced an “Outcomes” feature alongside Dreams, which allows agents to evaluate their own work against predefined metrics and retry tasks that fall short.
Source: mint – technology