Anthropic 'Dreaming' Feature Ships for Managed Agents: Scheduled Memory Curation That Improved Harvey's Task Completion 6x

VentureBeat·medium signal

Anthropic introduced 'dreaming' on May 6 as a research preview for Claude Managed Agents — a scheduled memory-curation process that reviews prior sessions, merges duplicates, removes outdated entries, and highlights recurring patterns between agent sessions. The system doesn't change model weights; it curates persistent memory and context for cleaner session starts. Harvey (legal AI) saw 6x improvement in task completion rates, and Wisedocs (medical document review) cut review time by 50%.