Agents and CodeHigh importance

Anthropic details agent memory and “dreaming”: Claude reorganizes logs in idle time

In the Engineering Blog (06/11/2026), Anthropic describes agent-memory components: memory notes, salience judgment and “dreaming” (sleep-time compute) to reorganize logs, correct errors and build representations (wikis). On CL-Bench 1.0, Opus 5 was the first of the cited series to close the learning loop with memory; Opus 5.5 and Sonnet 5.5 also. In Pokémon Red, Fable 5 wrote 9 files (4 badges) vs Sonnet 3.7 with 413 files (1 badge). On the Stainless corpus (~350k files / ~420M tokens), wikis after 128 dreaming sessions rose from 13% to 48% of criteria across 25 questions. The @AnthropicAI post had about 998,922 views, 9,364 likes, 808 reposts, 658 replies, 517 quotes and 7,884 bookmarks at the measurement (~08:12 BRT on 07/11).