Anthropic shares what it learned building memory for Claude agents

Anthropic's engineering team describes how Claude agents keep notes between sessions, plus a background step it calls dreaming, where Claude rereads past transcripts to fix wrong notes, build linked wiki pages and keep a small index loaded at the start of each session. In its tests, newer models did better when told only that a memory folder existed; a fixed rule to "record only failures" cut much of the gain for Opus 5 and Opus 4.7. On about 350,000 internal files, scores on 25 questions rose from 13% to 48% as wiki building grew from two to 128 sessions, and wiki plus search scored 61% against 53% for search alone while saving about 20,000 tokens per question. These are Anthropic's own evaluations, some on the game Pokemon Red, not independent tests. HN (75 comments): No consensus. One camp turns automatic memory off, deletes most memories Claude creates, and prefers AGENTS.md or progress files kept in the repo. Another dismisses the post as marketing and objects to the word dreaming. A smaller group found it useful, and one commenter read the charts as showing wiki plus search does most of the work. Sorted take: keep agent memory short, reviewed and in version control; the gains here came from curated notes, not from saving everything.