🧠Claude Learns to 'Dream' Its Way Through Memory Consolidation Between Sessions

Your agents could finally remember what they learned

Claude can now 'dream' its memory into shape. Between sessions, it replays logs, corrects its own mistakes, writes notes, and preloads what it expects to need next time. Opus 5 was the first model to close the loop on the Continual Learning Bench, using saved memory to store validated rules for answering future questions. If you're building agents, this is the part that matters. Most agent memory is a pile of transcripts nobody reads again. Claude now tags salient events as they happen and decides what's worth keeping, which means your agent stops re-deriving the same lessons every morning. The catch: memory is only as good as the judgment behind it, so you still need to audit what it writes. The benchmark setup is strict. Each CL-Bench question ran in a fresh session, with one exception: a memory directory that persisted across sessions. Claude's Pokemon runs show the spread. Sonnet 3.7 struggled because game state changes constantly, while Fable 5 stored facts the game never told it, and its global.md file worked as a dashboard with the current goal, a gym plan, and its own memory rules. Worth watching: whether these gains hold outside game and benchmark environments, where salience is messier.