Anthropic describes the components of agent memory and an idle-time “dreaming” step
6.3/10Significance
On November 6, 2026, Anthropic published an engineering post on agent memory. It says Opus 5.5 scored 0.838 with memory against 0.655 without on the CL-Bench database task.
What changed
Anthropic says dreaming raised performance on a 350,000-file wiki task from 13% to 48% of graded criteria over 128 sessions.
Why it matters
Memory that persists between sessions is a key part of making AI agents useful on long tasks.
What remains uncertain
Anthropic says dreaming requires compute and that one wiki grew larger than the documents it summarized. Gains were smaller for Sonnet 4.6, from 0.330 to 0.364.
What to watch
Watch for memory features in Claude products built on these ideas.
Source
- anthropic.comThe components of agent memory \ Anthropic
Why this ranks here
Useful primary research with measured results, but limited to Anthropic’s own tests.
- Impact6
- Reach6
- Novelty6
- Commitment5
- Coverage quality7
- Future potential7