Anthropic Explains How "Dreaming" Helps Claude Organize Its Memory
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.
Context
Memory that lasts between sessions is what lets an agent improve with experience instead of starting over, and Anthropic is showing how it builds it for Claude.
In an Engineering Blog post, Anthropic describes agent memory as knowledge Claude keeps between sessions, saved as context it can read later. It borrows "dreaming" from sleep-time compute, an idea from researchers at Letta and UC Berkeley: using idle time to reflect on session logs, fix mistakes and write new notes. The post explains an approach; it does not announce a new model.
Deep dive
Anthropic compares dreaming to how the brain is thought to consolidate memories during sleep, replaying recent experiences and keeping what matters. For Claude, that means reviewing logs between sessions, correcting mistakes and turning what it learned into notes it can find later.
The post walks through evaluations, including Continual Learning Bench (CL-Bench) 1.0, where Claude answers questions from a database in fresh sessions and can carry lessons forward only through a memory folder. According to Anthropic, models differ in how well they build their own learning loop: Claude Sonnet 4.6 got as far as generating theories about why an answer was wrong, leaving untested guesses in its memory files.
What to expect next
Watch whether dreaming becomes a feature in Claude or Claude Managed Agents, and whether Anthropic publishes memory results for newer models on CL-Bench.