Claude's agents now "dream" overnight: they re-read their own sessions and fix bad memories
Anthropic's November 6 engineering post explains dreaming: a background pass that reviews agent session transcripts and memory notes while the agent is idle, then reorganizes what future sessions will read. Dreaming can correct wrong notes against the original transcript, build wiki-style memory maps, promote high-signal rules into the context that loads at session start, and fill gaps a single session left behind. Anthropic also shipped dreaming as a research preview in Claude Managed Agents, where a dream reads a memory store plus past sessions and writes a separate cleaned store you can review before using. In one Pokémon run, fixing a single wrong landmark note with dreaming changed whether Claude escaped Mt. Moon.
Why it matters
Long-running agents fail when they trust stale or wrong notes. Dreaming is a concrete way to clean that memory offline instead of re-explaining the project every morning.
Who it's for
Teams building Claude Managed Agents, Claude Code power users who keep CLAUDE.md or auto-memory files, and anyone designing multi-session agent harnesses.
How to try it
- Read Anthropic's post at anthropic.com/research/components-of-agent-memory for how notes, wikis and context injection fit together.
- If you use Managed Agents, open the Dreams docs and request research-preview access, or adapt the cookbook dynamic-workflow recipe to run on a cron overnight.
- Point dreaming at a real memory folder plus 1–100 finished session transcripts, then review the new output store before mounting it read-only in the next session.
- Keep the injected top layer small (Claude Code only auto-loads the first ~200 lines / 25 KB of MEMORY.md) so promoted notes stay high-signal.
Plan & availability: Engineering write-up published November 6, 2026 on anthropic.com. Dreaming for Managed Agents is a research preview on the Claude Platform; developers request access. Cookbook and docs cover building your own dream workflow. Checked on Anthropic's post and Managed Agents dreams docs on November 7, 2026.
Source: @Anthropic on www.anthropic.com ↗
Written by AgentStackList from the original post. Checked Nov 7, 2026. Features, plans and availability can change, so confirm on the source before you rely on it.
agentsmemorydreamingmanaged-agentsclaude-code
More Claude tricks
- $200 free sandbox compute for your coding agents: Daytona, no card · Nov 7
- Free OpenFOAM CFD agent: describe a flow in plain English, Reynolds returns the case and plots · Nov 7
- 1,000 free web-search credits a month for your agents: Tavily Researcher, no card · Nov 7
- 1,251 physical designs in 12 hours: Archimedes agents rewrite hardware and re-check what fails · Nov 7
Agent Stack Weekly · free
Get the week's best AI tricks
Limit and pricing changes, new MCP servers, and setup fixes for Claude Code, Codex, Cursor, Copilot and Antigravity.
No spam. Unsubscribe anytime. Read past issues