Anthropic published “The components of agent memory” on 6 November 2026, describing evaluations of persistent memory and background consolidation it calls dreaming. The company reports that reviewing earlier transcripts can correct notes and organize knowledge for later sessions, while acknowledging that memory can accumulate errors and become outdated.[1]
The article defines agent memory as knowledge retained between sessions in readable context, excluding model weight changes and conversation compaction. Dreaming revisits logs to revise notes, build linked wikis or select material for automatic loading into future sessions. Anthropic warns that an incorrect injected note can recur in every session, and loading a note does not ensure Claude follows it.[1]
In an evaluation using internal documents from Stainless, Anthropic reports that combining search with an expanding wiki raised aggregate scores from 54% to 61%, compared with 53% for search alone, across 25 questions. Results varied by task: one legal-brief evaluation showed no clear gain from a wiki, while another produced a wiki larger than its source documents.[1]
A contract exercise exposed a separate failure: agents overwrote one another’s contributions to a shared table. Anthropic reports that revised instructions assigned each contract to one agent and one page, retaining verbatim terms; the resulting brief could meet all graded criteria, compared with 72% previously. These are company-reported evaluation outcomes, not independently reproduced results.[1]
[1]: Anthropic, “The components of agent memory,” 6 November 2026. https://www.anthropic.com/research/components-of-agent-memory
Sources
Check or correct this page
Verification record
The digest below shows whether the text you are reading matches the published source. It does not show that the reporting is correct. The build computes it from this page's Markdown source and lists it in the site manifest.
3181cab060fb9f529a12f2444761d3c4bc7ef5ecb1029e27ab4c80419d05957f