Anthropic Lets Claude Dream Between Sessions

🔥 What happened Anthropic unveiled "dreaming": Claude reorganizes its own logs between sessions, fixes mistakes, and writes new memory notes. The design borrows from hippocampal replay research and Letta/UC Berkeley's "sleep-time compute" concept. 💡 Why it matters Agent memory shifts from passive storage to an active learning loop. On CL-Bench, only Opus 5 fully closed the loop — earlier models just piled up untested guesses. The gap between "taking notes" and "validating notes" is what separates long-horizon competence from confident drift. ⚡ Our take Finally someone treats agent memory as architecture, not a prompt hack. But who audits Claude's self-authored rules before they harden into elegant wrongness?