Anthropic enables Claude to "dream" to organize its memory, with 350,000 documents, boosting its question answering score from 53% to 61%.

Beating AI Express: Anthropic Publishes Research on Agent Long-Term Memory, Details Claude’s "Dreaming" Mechanism After agents complete ongoing tasks, they accumulate large volumes of conversations, operational logs, and notes. However, this data can become outdated, contradictory, or hard to retrieve when needed. Anthropic divides memory into four layers: raw records, curated notes, an interconnected Wiki knowledge base, and a concise index that loads automatically each time the model starts. Claude can jot down key information during task execution. Its "dreaming" mechanism reviews historical records between tasks, correcting errors, deleting outdated content, and distilling new insights from multiple task runs. On subsequent launches, Claude only needs to load the concise index and retrieve relevant memories as needed, rather than re-reading all historical data. Anthropic conducted experiments using approximately 350,000 internal documents from Stainless, totaling 420 million tokens. Researchers had Opus 5 complete 128 rounds of sessions, during which it continuously read documents, built the Wiki, merged duplicate content, and verified facts. Across 25 questions about these documents, the model scored 53% when using only a search tool, rising to 61% after integrating the evolving Wiki. It also saved an average of roughly 20,000 tokens per question. The Wiki’s advantage was particularly pronounced for questions requiring synthesis of multiple contracts; for answers contained in a single document, standard search sufficed. Anthropic also tested Claude on *Pokémon Red*. The older Sonnet 3.7 model logged 413 memory files but earned only one badge. The new Fable 5 model retained just 10 files yet secured four badges, demonstrating the new model’s improved ability to distinguish between strategies worth preserving and game states that can be re-accessed as needed.