Claude Managed Agents at a glance
Anthropic published 'The components of agent memory' on November 6, 2026, detailing Claude dreaming: an offline pass in which Claude re-reads past session transcripts to correct its memory notes, merge duplicates and build wikis; one dream fixed a wrong note that had trapped Claude in Pokémon Red.
What changed with Claude Managed Agents
Anthropic published 'The components of agent memory' on November 6, 2026, a research post that defines Claude dreaming and the memory system behind Claude Managed Agents and its Dreams feature. Dreaming is an offline pass, run while an agent is idle, in which Claude re-reads its memory notes against the original session transcripts, corrects mistakes, merges duplicates, builds structured representations such as wikis, and decides what future sessions should load first.
Anthropic tested what models choose to remember in Pokémon Red, comparing memory folders at about 10,000 steps into the game: Claude Sonnet 3.7 wrote 413 memory files across 225 folders, 977,930 characters in total, and earned one badge, while Claude Fable 5 kept 10 files totaling 43,406 characters and earned four badges. In the same game, Claude had wrongly recorded that the PokéCenter stood next to the PokéMart in Pallet Town, which had made Claude keep going back there to heal — dreaming corrected the entry and Claude completed the game.
Dreaming also builds new representations: working from an internal corpus of about 350,000 files, roughly 20 percent of the Anthropic codebase, an agent built a hierarchical 5,000-file wiki in 20 hours, making 2,000 changes per hour while humans made 10; a separate experiment built an internal agent that uses dreams to keep memory current, called 'an organization that remembers' in the post's introduction, and Anthropic says dreaming reduced incorrect facts by about 90 percent in one of its evaluations, with more detail in the post's section on evaluating dreaming.
For developers, the same mechanism ships as Dreams in Claude Managed Agents, documented by Anthropic as an agent-run offline pass that merges duplicates, fixes mistakes and organizes the memory files stored in the agent's .claude directory, and Anthropic lists Claude Managed Agents as in public beta.
Why Claude Managed Agents matters for businesses
Why it matters: an agent that keeps its own memory accurate gets more dependable the longer it works, which matters for customer work, support and long-running projects, and Claude Managed Agents targets exactly that long-running use — but the research case still needs confirming in production: a test that works in a game does not prove the same gains in live business workflows, so the check is whether memory accuracy holds up over weeks of real work.
Sources
- The components of agent memory — Anthropic · anthropic.com · Official
- Dreams — Claude Platform Docs (Managed Agents) — Anthropic · platform.claude.com · Official
Last verified: November 21, 2026. IA Compa summarizes public announcements for business readers. We are not affiliated with the companies in these stories.
Top comments (0)