Watch the video, transcript and sources
https://www.youtube.com/watch?v=psdH4QXD0Us
- A wrong memory repeats the same mistake
+ Transcript-backed review repairs the note
+ Separate contract pages preserve exact termsGiving an agent a memory can preserve its mistakes across sessions. In Anthropic's new components of agent memory article, a wrong landmark sent Claude over the same Pokémon ledge three times. “Dreaming” is the company's name for extra model sessions that revisit transcripts and reorganize stored context. The model's weights stay unchanged.
The setup matters. Claude wrote that a Pokémon Center was on the east side of Mount Moon, although the building was near the west entrance. It then left believing the cave was behind it. A one-way ledge dropped it onto the earlier route, and its navigation tool missed the drop. Claude explained the failure as overshooting, preserving the wrong assumption for another attempt.
After the third fall, it wrote that it was going to stop rather than keep looping. The researchers' automatic reply told it to continue. The uncorrected run ended without reaching Cerulean City. A separate dreaming pass checked the wrong note against the original transcript, corrected it, and the corrected branch eventually reached that city. This is Anthropic's reported experiment; we inspected its public source rather than running the game ourselves.
A wiki helps some questions
Anthropic also organized Stainless's internal documents into a linked wiki. Search alone scored 53% of the graded criteria; search with the larger wiki reached 61%. The evaluation contained 25 questions about that corpus. The article doesn't state the total cost of building the wiki, while the Dreams documentation bills maintenance at standard model API rates.
The limits are unusually useful. One legal benchmark task showed no clear wiki gain over search. In another, the wiki grew larger than the documents it summarized. A question answered by finding one passage may have little use for an encyclopedia built in advance.
The shared-table bug
The practical payoff comes from a simulated acquisition. Several agents wrote contract terms into one shared table. Summaries lost important terms, and writers overwrote one another's work. The resulting brief included 41 of 57 graded facts.
Revised wiki instructions gave each contract a separate page with a single writer. The terms were quoted word for word. The improved wiki supported all 57 graded facts. Ownership and quotation rules changed together, so the test doesn't isolate either change's contribution. These are benchmark criteria, with no claim of perfect real-world legal advice.
Verdict: NEEDS REVIEW
I'd use transcript-backed repair, inspect the changed notes and test whether the wiki improves my actual task. Managed Agents' Dreams preserves the input store and produces a separate output that can be reviewed. Its research-preview status also means this article doesn't activate dreaming for every Claude user.
The custom-workflow cookbook warns that memory is only as good as its transcripts. Yesterday's log cannot establish today's changed facts. Important notes need a path back to their sources and a reason to remain in the next session.
Sources and media credits
Anthropic research — Lance Martin, figures by Monika Tuchowska
Independent earlier sleep-time compute research, April 2025 — prior work, no replication of this article's results
Portrait: Dario Amodei at TechCrunch Disrupt 2023, TechCrunch, CC BY 2.0; cropped, background removed and resized. Anthropic identity only; no authorship, current reaction or endorsement implied.
Two Buttons: Jake Clark, origin attribution via Supermeme, with commentary labels. Groundhog Day GIF: Bill Murray/Stephen Tobolowsky as Phil/Ned, @groundhogday; silent, resized and captioned. Both are brief transformed narrator commentary, with no present-day reaction by anyone shown. Hosting is not a licence; no legal-clearance or blanket-reuse claim.
And that's the diff for today. I'm Niko from Axrisi. Merge responsibly.