Anthropic shares its agent memory tests, from notes to offline “dreaming”
TL;DR
- Anthropic published “The components of agent memory” on 6 November, on how Claude writes notes it keeps between sessions and how “dreaming” rereads transcripts to fix and organise them.
- On Anthropic's own 25-question test over about 350,000 internal files, scaling dreaming from two sessions to 128 raised Claude's wiki-only score from 13% to 48% of graded criteria.
- A rule to “record only failures” cost Opus 5 and Opus 4.7 much of their gain from memory.
Read the full story
Sign in with your email to read AI News. It’s free.