Imagine you're moving into a new apartment and you tape a note to every single light switch: 'This turns on the kitchen light.' 'This turns on the bathroom fan.' Within a week your walls are covered in tape and you can't find the one note that actually matters — the circuit breaker location for when the power trips. That's the difference between Claude Sonnet 3.7's memory strategy (413 files, one Pokémon badge) and Claude Fable 5's approach (nine files, four badges). The core mechanism here isn't about storage capacity. It's about salience filtering — predicting which experiences will matter in the future and discarding the rest. Anthropic's post frames agent memory as a three-part system: real-time note-taking during sessions, offline 'dreaming' that consolidates and corrects those notes between sessions, and salience judgment that determines what's worth writing down at all. The biological analogy is explicit and structural: the hippocampus captures experiences fast, sleep replays and distills them, and chemical tagging marks what matters. The engineering claim is that Claude's frontier models have learned to close this loop without being told how. The strongest evidence comes from two evaluations. On CL-Bench 1.0, a benchmark for long-horizon learning, the progression across model generations is stark. Sonnet 4.6 generated untested guesses. Opus 4.7 consolidated notes and tracked confirmation status. Sonnet 5 auto-corrected on failure. Opus 5 closed the full learning loop — writing rules, citing evidence for those rules, and flagging traps. With memory enabled, Opus 5 through Sonnet 5.5 all showed measurable performance gains over their no-memory baselines. The Pokémon Red evaluation illustrates the salience problem most vividly. Sonnet 3.7 treated memory as an archive, recording every battle, item, and NPC encounter in separate files. By step 10,000 it was spending more compute managing its memory than playing the game — creating six files and deleting nine others across 17 steps with zero in-game actions. Fable 5, by contrast, kept a single dashboard file (global.md) with its current goal, gym strategy, and self-imposed memory rules. Its memory stored only things the game wouldn't tell it: strategies, plans, and hard-won rules like varying Poké Ball timing after three consecutive failures. The steering results add nuance. Telling weaker models to read their memory files before answering nearly doubled Sonnet 4.6's CL-Bench score. But for stronger models, prescriptive rules hurt. When Opus 5 was told to 'record only failures,' it lost much of the performance boost that unconstrained memory provided. The practical takeaway: tell Claude what memory is for, not how to use it. Frontier models already judge salience better than fixed heuristics. Dreaming — offline reflection on session transcripts — serves as error correction. Claude reviews its own notes against the original logs and can identify drift or outright mistakes. The post frames this as analogous to sleep replay in the brain, where the hippocampus replays recent experiences to consolidate them into long-term cortical storage. The engineering implementation borrows from Letta and UC Berkeley's 'sleep-time compute' concept: using idle time between interactions to reorganize existing context. What's missing is as instructive as what's present. There are no ablation studies isolating dreaming's contribution from simple note-taking. The Pokémon evaluation compares models of different capability levels rather than the same model with and without dreaming. The CL-Bench results show memory helping, but the dreaming component's marginal contribution isn't quantified separately. This is an engineering blog post, not a paper — and it reads like one. The insights are real but the validation is internal, uncontrolled, and selected for narrative clarity.
