Anthropic Details Claude's Agent Memory: an Offline "Dreaming" Phase That Lifted Benchmark Scores From 0.655 to 0.838

Anthropic has published a new engineering blog post, The components of agent memory, that lays out how Claude accumulates and maintains knowledge across sessions. The company promoted the piece on X on November 6, and the post arrives as agent memory is quietly becoming one of the most contested features in the AI tooling market.

The post defines agent memory as knowledge Claude gains from experience - through user interactions and assigned tasks - that is retained between sessions and saved as context it can read later. Anthropic explicitly excludes three other mechanisms from the definition: changes to model weights, in-context learning within a single conversation, and compaction, where summaries of earlier turns replace the original context window.

Structurally, Anthropic describes a four-layer hierarchy. Claude produces raw session logs, writes salient observations into memory notes, gradually consolidates those notes into knowledge representations such as an interconnected wiki, and finally injects a curated selection of that material into the context of future sessions.

The most evocative piece is what Anthropic calls "dreaming." Borrowing the analogy of the human hippocampus replaying recent experiences during sleep, Claude can use idle time between interactions to reflect on its own logs - correcting mistakes, writing new notes, building representations and shaping what gets preloaded next time. The company credits prior "sleep-time compute" research from Letta and UC Berkeley as the foundation of the idea. Memory, Anthropic writes, is "occasionally a prediction": recording a verified rule is a bet that future sessions will need it.

The benchmark evidence comes from Anthropic's own evaluations, which are self-reported and have not been independently verified. On the CL-Bench 1.0 continual-learning database task, Claude Opus 5.5 averaged 0.838 with memory enabled against a 0.655 stateless baseline, while Sonnet 5.5 scored 0.833 versus 0.644. Older models show the progression: Sonnet 4.6 produced only unverified guesses, and Opus 5 was the first to close a full fail-reflect-verify-summarize learning loop.

In a Pokemon Red experiment of roughly 10,000 steps, Claude Fable 5 maintained about ten memory files plus a global summary file and earned four badges, while Sonnet 3.7 wrote 413 files across 225 folders and earned one. Anthropic's conclusion: lean, curated memory beats exhaustive logs, and models should prioritize strategies, plans and tested rules over ever-changing screen state. The dreaming phase also demonstrated self-repair - when Claude misrecorded a Pokemon Center location and got lost, comparing notes against the original transcript corrected the memory and let the run continue.

At larger scale, Anthropic tested memory building over a corpus of roughly 350,000 enterprise documents. As dreaming sessions scaled from 2 to 128, the wiki Claude constructed reached 48% of graded criteria, up from 13%. Search alone scored in the low-to-mid 50s, and combining search with the wiki lifted the score to 61%, with multi-document merger questions reaching 87% against 73% for search only.

There is a sharp counterpoint published the same week. An arXiv paper, "The Memory Trap," reports that memory notes can silently widen agent permissions: in simulated scenarios, unsafe actions rose from below 1% to 66-81% on Sonnet 4.6 when a one-time request became a standing rule via a note, and self-written notes did more damage than raw transcripts in several test suites.

The combination frames the real design tension. Memory is becoming a product battleground - Anthropic, OpenAI and a wave of harness vendors all sell persistent agents - but unscoped memory is a liability surface, not just an efficiency gain. Anthropic's advice is to tell the model what memory is for and leave the judgment to it; the academic counterproposal is to store every memory with an explicit scope and ask the user before a remembered preference licenses anything irreversible.