Anthropic explains how letting Claude "dream" improves its long-term memory

Anthropic published research on November 6 describing a technique it calls dreaming, where Claude uses idle time to reorganize its own memory notes and build linked wiki-style pages, without any changes to the underlying model weights. The idea is similar to how people consolidate memories during rest: the agent revisits what it has learned and restructures it for easier recall later. Anthropic tested this on a large corpus of Stainless company files, about 350,000 documents and roughly 420 million tokens. When they scaled the dreaming process from 2 sessions up to 128 sessions, performance on a set of 25 test questions jumped from 13% to 48% of graded criteria met. Combining simple search with the new wiki system also helped: the aggregate score rose from 53% with search alone to 61% when the wiki was added. Beyond accuracy, the approach was also more efficient, saving around 20,000 tokens per question compared to not using it. This suggests that giving AI agents a structured way to revisit and reorganize what they've learned can make them both smarter and cheaper to run over long-term tasks.

Why it matters

Better memory organization without retraining could make long-running AI agents noticeably more accurate and cheaper to operate.

Source: Anthropic