Anthropic tests Claude dreaming, an idle-time process that audits its memories
Anthropic's engineering team has described a mechanism it calls dreaming, which lets Claude review earlier sessions while it sits idle. Interesting Engineering reports that Claude compares stored memory notes with the original conversation transcripts, flags inaccurate entries, refines them, and organizes what it has learned for later sessions. The write-up, dated 6 November 2026, says the design takes loose inspiration from human sleep, though the mechanism differs.
The test used Pokémon Red, where the agent had to explore the game world and collect badges. Researchers compared two models after roughly 10,000 gameplay steps. In one case, an incorrect note placed a Pokémon Center on the wrong side of Mt. Moon, sending the agent off a one-way ledge. After dreaming checked that note against the transcript, Claude fixed it, re-entered the cave and reached Cerulean City. The core tradeoff: logging every step clutters memory, while leaving out details starves later sessions of context.
Other reports add figures. Sonnet 3.7 saved 413 memory files and earned one badge, while a newer model called Fable 5 saved nine and earned four, though one report counts ten files. In a wiki test on about 350,000 internal documents, the score rose from 13% to 48% as dreaming sessions grew from two to 128. Several results come from single runs.
Why it matters
For teams building long-running agents, a wrong memory note does not stay contained. It shapes every later session that reads it, so errors compound. A periodic audit against raw logs is cheap to test and can run before a team adds more context to an agent. Teams should also keep injected memory small, since Claude Code reportedly loads only the first 200 lines or 25 KB of its memory index into each session.
For founders and investors, the change sits in the external memory layer rather than the model weights, so memory-driven products can improve without retraining. Anthropic itself notes that dreaming requires compute, so audit loops carry a running cost that should appear in pricing and margin models. Governance matters too: edits to memory that affect customers need provenance, review and deletion rules before they reach production.
Key points
- 1Dreaming runs while Claude is idle and checks memory notes against the original session transcripts.
- 2The Pokémon Red comparison used one run per model at about 10,000 gameplay steps.
- 3A corrected note about a Pokémon Center let the agent reach Cerulean City after a navigation loop.
- 4In a wiki test on about 350,000 documents, graded scores rose from 13% to 48% with more dreaming sessions.
- 5Anthropic says dreaming consumes compute, so teams should aim it at information likely to change.
