Anthropic studies Claude's cross-session memory mechanism
Understand this story
Anthropic published research stating that the key to Claude's cross-session memory is selective recording rather than full archiving. In a Pokémon Red test, Claude Sonnet 3.7 generated 413 files and 225 folders in about 10,000 steps but ultimately earned only 1 badge; Claude Fable 5 recorded only strategies and verified rules and earned 4 badges. The research also introduced the Dreaming correction mechanism, used to retain knowledge across sessions while the model remains unchanged. Dreaming was launched in May as a research preview of Claude Managed Agents and initially supported only Claude Opus 4.7 and Sonnet 4.6.
AI overview based on source reports · Updated 2 hr ago
Coverage timeline
Follow the coverage to explore different perspectives.
- The DecoderAnthropic says the key to agent memory is knowing what to forget, not how much to store
Anthropic research shows that Claude agents retaining knowledge across sessions perform better with selective notes than exhaustive archives: in a Pokémon Red test, Claude Sonnet 3.7 created 413 files across 225 folders and earned one badge, while Claude Fable 5 kept a few files and earned four. Anthropic's "Dreaming" feature has Claude review transcripts of past sessions to correct faulty notes, which fixed an error that had trapped Claude three times, and it built a linked wiki from about 350,000 Stainless documents, scoring 87 percent on a legal test versus 73 percent with search alone. Anthropic notes open issues including distraction from note-taking, extra compute cost, and the need for regular upkeep.
Activity for this story
Not enough continuous observations to show a trend yet.