Anthropic's Claude begins 'dreaming' to reorganize memory in tests
Anthropic announced that its Claude model can enter a post‑dialogue "dream" phase, using idle compute to replay the day’s interaction logs. In these dreams the model corrects basic errors, extracts strategy tips and builds a pre‑loaded Wiki that can be consulted in future sessions. The blog notes that the idea is separate from updating model weights and that any link to the upcoming "Model 4" or…
Key points
- Claude uses "dream" sessions to review logs, correct errors and build an external Wiki
- Increasing dream sessions from 2 to 128 on 350,000 files raised Wiki‑only scoring from 13% to 48% (61% with search)
- In a Pokémon Red test, memory‑efficient Fable 5 wrote 9 files and earned 4 badges, versus Sonnet 3.7’s 413 files and 1 badge
The company demonstrated the approach with two experiments. In a Pokémon Red playthrough, the memory‑lean Fable 5 wrote only nine files (plus a global note) and earned four badges, while the older Sonnet 3.7 logged 413 files across 225 folders and secured just one badge. A separate corporate test fed Claude 350,000 internal documents (about 420 million tokens) from the software firm Stainless. Raising the number of dream sessions from two to 128 lifted the Wiki‑only scoring‑criteria satisfaction rate from 13% to 48%, and to 61% when combined with search tools. The work draws on the "sleep‑time compute" concept proposed by researchers at UC Berkeley.
Claude has started to 'dream'! How does 'Dreaming' change the memory of a forgetful AI?
note.com · 6 November 2026
Loading the full article…
This text was published by note.com and written by Trans-N. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Research
All →- John Langford quits Microsoft, declares AI Frontiers over · 1 src
- MIT and BASF release CheMeleon for low-data molecular prediction · 2 src
- OpenAI pipeline finds 948 Linux privilege escalation flaws using upcoming cyber model · 11 src
- Bartowski releases Xing4.0-29B-A4B quantized models for llama.cpp · 1 src
- Anthropic announces FlashProve to prove AI answers · 1 src
Comments
via GitHub Discussions