Anthropic says letting Claude 'dream' between tasks helps it fix and organize its own memory
In company tests, the review step caught a wrong note that had kept Claude lost in a video game. Businesses can try it in a test version.
Anthropic, the San Francisco company that makes the Claude AI system, says a process it calls dreaming helps its AI agents keep better memories from one job to the next, according to a research post it published Nov. 6. An AI agent is an AI that carries out multistep tasks on its own instead of just answering one question.
Why AI needs notes
A chatbot normally forgets everything when a conversation ends. To get around that, agents can write notes into files and read them in later sessions, much like a worker keeping a notebook. The problem, Anthropic says, is that a note can be wrong from the start or go out of date, and the agent will still trust it later.
Dreaming borrows an idea from how human brains are thought to sort memories during sleep. When Claude is idle, it rereads the full records of its earlier sessions, checks its notes against them, fixes mistakes, merges duplicates and builds linked summary pages, like a small personal Wikipedia.
What the tests showed
In one test, Claude played the 1990s video game Pokémon Red and wrote down the wrong side of a mountain for a landmark. Without a fix, it kept walking off a ledge and never reached the next town. After dreaming caught and rewrote the note, it went back into the mountain and made it through, the company says. In another test using about 350,000 internal files from Stainless, a software company Anthropic bought in May, Claude's score on 25 questions rose from 13% to 48% as it built and improved its summary pages over 128 sessions.
Anthropic also found that its newer models do better when left to judge for themselves what is worth remembering, rather than following fixed rules. It says dreaming takes extra computing power. In the same research post, Anthropic says several questions about agent memory remain open: when Claude should write memory during a live task versus later in a dream; how to keep notes reliably up to date as the world changes without spending too much computing power; and whether searching raw session transcripts should be the foundation of memory, with tools such as wikis added on top. Software developers can try it through Claude's business platform as a research preview, an early test version.
Source: Anthropic research post; Claude developer documentation on dreams