Anthropic details Dreaming memory technique for Claude

Anthropic published the working details and experimental results of Dreaming, its memory-review technique for Claude, on its official blog on November 6, 2026. According to aitimes.com, the technique tasks AI agents with reviewing past work logs and correcting their own faulty memories.

In one test, Claude Sonnet 3.7 played Pokémon Red, creating 413 files in 225 folders over roughly 10,000 steps and earning one badge. Claude Fable 5 earned four badges while keeping two folders, nine files and one summary file. With Dreaming, Claude corrected a wrong memory of a Pokémon Center location and reached the next town, while a version using the uncorrected memory did not.

Anthropic also had Claude Opus 5 build a wiki-style knowledge base from about 350,000 internal documents totaling 420 million tokens across more than 128 analysis passes. Wiki-only evaluation scores rose from 13% to 48%, search-only scored 53%, and combining search with the wiki lifted results from 54% to 61%. Combining both cut token use by about 20,000 per question, and on synthesis-heavy questions such as M&A queries the wiki approach scored 87% against 73% for search alone.

In a separate contract-analysis test, Claude changed its memory rules so each contract received its own document, raising compliance from 72% to 100%.

Anthropic reported limits as well: Dreaming did not improve some legal document analyses, wikis sometimes grew larger than the source material, and reviewing memories requires additional compute.

Earlier memory efforts

Anthropic first introduced Dreaming in May 2025. In June 2025, Perplexity released Brain for analyzing agent work experience, and OpenAI strengthened ChatGPT's memory system in the same month.