Anthropic has published experiments on memory that AI agents retain between sessions. The work examines correcting notes against earlier conversations and organizing document collections into linked pages. These company-run tests distinguish remembering useful information from repeatedly carrying mistakes forward, with results depending on the task and the way information is organized.
Anthropic, the developer of Claude, has published its own experiments on information agents keep between sessions. These memories are files and notes a later session can read. In a Pokémon game experiment, an incorrect location note repeatedly sent an agent the wrong way. Comparing that note with the original conversation allowed the agent to correct it before continuing. The company calls this background reorganization process dreaming.[1]
Other experiments organize information into a wiki: linked pages that summarize a subject and point to related notes. One collection comprises about 350,000 files from software company Stainless, tested through 25 questions. Anthropic compares direct search with these pages and with the two methods combined. Legal-document tasks distinguish questions needing information from several documents from answers readily available in a single file.[1]
Agents editing a shared contract table sometimes overwrote information. Giving each contract its own page was one response to the loss of information. On another task, the generated wiki became longer than the documents it represented. The company also identifies the danger of preloading incorrect notes into every new session. These examples examine memory maintenance in controlled tasks, rather than announcing a new consumer feature or independently establishing general reliability.[1]
