Science fiction has long imagined machines that think, learn, and perhaps even dream. Anthropic is now exploring an idea that sounds straight out of that world. Its latest research examines how Claude can revisit past experiences during periods of inactivity, correct faulty memories, and organize what it has learned for future conversations. Anthropic calls this process “dreaming,” although it has little to do with the dreams humans experience.The approach allows Claude to review previous interactions when it has idle time. It can identify incorrect information, refine existing notes, and build structured knowledge that future sessions can access. The process draws inspiration from how human brains consolidate memories during sleep.Anthropic’s research examines several components of this system. Its experiments suggest that selective memory and regular maintenance can improve performance on complex, long-running tasks.Teaching ClaudeClaude can save useful information in memory files while completing tasks. However, deciding what deserves storage presents a challenge. Recording every interaction creates clutter, while overlooking important details can leave future sessions without useful context.Anthropic tested this problem using Pokémon Red, where Claude had to navigate the game’s world and earn badges. The researchers compared two models after approximately 10,000 gameplay steps.Memory is what allows Claude to learn from experience. In a new Engineering Blog post, we share how “dreaming” helps Claude organize what it’s learned, correct its mistakes, and find what it needs more easily the next time. https://t.co/uUEN4bNkr9— Anthropic (@AnthropicAI) November 6, 2026Sonnet 3.7 recorded almost everything it encountered. Its memory directory contained 413 files across 225 folders, totaling nearly 978,000 characters. Despite maintaining this enormous record, the model earned just one badge.Claude Fable 5 took a considerably different approach. It maintained only 10 files across three locations, using approximately 43,400 characters. The model earned four badges during its run.The smaller memory system prioritized navigation strategies and lessons learned through gameplay. It also maintained a central file containing its current objectives and planned actions.This approach helped Claude avoid repeatedly processing information that it could already retrieve from the game’s interface. It also preserved useful discoveries that would otherwise disappear between sessions.Dreaming helps correct mistakesMemory files can contain incorrect information, creating problems when future sessions treat those details as established facts. Anthropic demonstrated how dreaming could address this issue through another Pokémon experiment.Claude incorrectly recorded the location of a Pokémon Center near Mt. Moon. The mistake caused it to misunderstand its position after leaving the building. It repeatedly walked over a one-way ledge and returned to an earlier route.An offline review changed the outcome. Claude compared its stored notes against the original session transcript and identified the incorrect location. It then updated the memory file with the correct information.With that correction, Claude re-entered Mt. Moon and eventually reached Cerulean City. The version using incorrect memory repeatedly encountered the same obstacle and never completed the journey.Building knowledge beyond conversationsAnthropic also explored whether dreaming could turn large document collections into interconnected knowledge bases. These systems resemble wikis, with individual pages linking related information.One experiment involved approximately 350,000 internal documents from software company Stainless. The collection contained roughly 420 million tokens.Across 128 sessions, Claude gradually developed and refined a wiki using information from those documents. Its performance on 25 evaluation questions improved from 13 percent to 48 percent.A separate merger-document experiment produced another notable result. Claude answered questions requiring information from multiple documents with 87 percent accuracy using its wiki. Direct document search alone achieved 73 percent.Dreaming can also reorganize information for future sessions. Claude can move frequently needed details into readily available context while keeping less relevant information in searchable files.Anthropic acknowledges that the approach still requires careful management. Memory maintenance consumes computing resources, and excessive information can overwhelm future sessions.The researchers believe agents must learn when to preserve information and when to revisit it. Their experiments suggest that better memory management could help Claude handle increasingly complicated tasks across multiple sessions.