We all know Anthropic has been struggling with one core issue in Claude Code: memory consistency across sessions. Auto-memory was a big step because it lets Claude take notes on your project and persist context over time, but after 20+ sessions that memory starts to decay. You end up with messy, conflicting notes, outdated debugging steps, and vague timestamps like “yesterday” that turn into pure noise—then hallucinations start creeping in. In this video I break down the quiet upgrade that basically fixes this: Memory 2.0 via a new system called Auto Dream (aka “dreaming”). Think of it like REM sleep for your AI agent. It consolidates memory between sessions by scanning your memory files, gathering high-signal corrections/decisions, consolidating (merging duplicates, resolving contradictions, converting fake timestamps into real dates), then pruning and indexing so startup memory stays short and useful. I also show how to enable it inside /memory (Auto Dream toggle), how to manually trigger it by prompting it to “consolidate my memory using dream,” and when you should run it (big refactors, framework switches, lots of trial-and-error debugging). The end result is fewer repeated outdated solutions, fewer hallucinations, and a cleaner per-project mental model that actually improves over time.