Anthropic on agent memory and ‘dreaming’
Confirmed
The takeaway
An Anthropic Engineering post (6 Nov 2026, Lance Martin) describes “dreaming” in Claude Managed Agents, where between sessions Claude reviews logs, corrects notes and builds wikis.
Status
- Claim
Confirmed
- Our reporting
- High confidence
- Importance
- 2 of 5
- Last verified
- 8 November 2026
Your AI and this story
- GPT-6 Astra190 days after its cutoff
- Claude Opus 5.5129 days after its cutoff
- Gemini 4 Argon220 days after its cutoff
- Grok 4.7159 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 129 days before it.
Key facts
- Anthropic’s X announcement reached ~1.77M views by Nov 8
- CL-Bench (baseline → with memory): Opus 5.5 0.655 → 0.838; Opus 5 0.564 → 0.828; Opus 4.7 0.383 → 0.700; Sonnet 4.6 0.330 → 0.364
- Pokémon Red: Fable 5 kept 9 files and earned four badges, against Sonnet 3.7’s 413 files and one badge; a dream-corrected landmark note got the agent past Mt. Moon
- Docs wiki on ~350k files: wiki-only accuracy 13% → 48% over 128 dreaming sessions; wiki+search 61% vs 53% search alone, ~20k tokens saved per question
What happened
This is an engineering explainer with experiments, not a new launch. Dreaming already exists in Claude Managed Agents.
Why it matters
It gives concrete numbers on how much learned memory between sessions helps current models.
Sources
2 sources from 2 sites. Numbers match the chips in the text.
2 sources: 2 primary
Primary
- Anthropic on X: dreamingx.com, official
- Anthropic: Components of agent memoryanthropic.com, official
Changes
- X post
- Filed (lab X list)