Anthropic details Claude’s “dreaming” memory architecture
Anthropic described a Claude agent-memory system that uses idle time between interactions to review session logs, correct erroneous notes, merge duplicate information, and build structured knowledge for later sessions. The architecture stores memories externally in files, notes, and wikis rather than changing model weights. Anthropic reported improved results in CL-Bench, Pokémon Red, and a test involving about 350,000 enterprise documents, including better memory-assisted scores and correction of a navigation error. These results were reported by Anthropic.
The work presents a concrete approach to persistent agent memory: use idle computation to reorganize external representations and correct errors before they affect later sessions. The reports also identify compute costs, outdated knowledge, and the lack of a standard dreaming schedule as unresolved design issues.
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://www.anthropic.com/research/components-of-agent-memory
#Anthropic#Claude#Claude Opus 5.5#Claude Sonnet 5.5#Claude Fable 5#Claude Sonnet 3.7#CL-Bench 1.0#Managed Agents