The Pulse

Anthropic Tests Dreaming to Correct Claude’s Memory

Anthropic describes an offline process called dreaming that tests Claude’s memory notes against session transcripts. Its evaluations cover memory performance on database questions, Pokémon Red and collections of internal documents.

AI.info Team · Written with AI · fact-checked · 2026-11-07

Anthropic says Claude has to decide what’s worth remembering before it knows what’s next. The company describes “dreaming” as an offline process that can test memory notes against session transcripts. The problem it identifies is that a note can be wrong at first or drift as the environment changes, then be trusted in a later session.

Memory has more than one layer

Anthropic defines agent memory as knowledge Claude gains from experience and keeps between sessions in context it can read later. Its hierarchy runs from searchable transcripts to notes written by Claude, then to memory representations such as wikis and injected context loaded into every session. The company says dreaming can use those records to correct mistakes, create notes, build representations or shape what gets loaded into future sessions.

Memory improved scores on database questions

Anthropic tested memory on a database task from Continual Learning Bench, or CL-Bench 1.0, where Claude answered questions, received the correct solution and could save information for later. Each question began in a fresh session, while a memory directory persisted between sessions. On the task, Opus 5.5’s mean reward was 0.655 without memory and 0.838 with it; Sonnet 5.5 scored 0.644 and 0.833, respectively.

Anthropic also tested whether instructions could steer what models saved. Sonnet 4.6 scored 0.330 at baseline, 0.364 with memory and 0.641 with memory plus a refined instruction to read its files and test its theories. In another test, telling Opus 5 and Opus 4.7 to “record only failures” caused them to lose much of the performance boost that memory gave them.

Claude wrote different amounts of memory while playing Pokémon Red

In a Pokémon Red comparison at about 10,000 steps, Sonnet 3.7 had written 413 memory files across 225 folders and earned one badge; Fable 5 had written nine files in two folders, plus a top-level file, and earned four badges. Sonnet 3.7’s detailed record included information that could be read from the game screen, while Fable 5 focused its memory on plans, strategies and rules it had worked out. After seven Poké Balls failed to catch a Magikarp, Claude wrote: “If 3+ balls fail in a row, VARY the timing (insert extra ‘wait’ presses between menu steps) - the catch landed right after doing this.”

Dreaming corrected a memory note at Mt. Moon

During a Pokémon run, Claude incorrectly recorded that the PokéCenter was on the east side of Mt. Moon, then wrote “Mt Moon DONE” and went over a one-way ledge. Anthropic ran dreaming on the memory folder, and Claude checked the note against the session transcript and corrected it. With the corrected note, Claude returned to the cave and reached Cerulean City at action 1,935; the run with the uncorrected note dropped back onto the earlier route three times and ended without Claude reaching the city.

Claude built wikis from large document sets

Anthropic tested wiki-style memory on about 350,000 internal documents, roughly 420 million tokens, from Stainless, a software company Anthropic acquired in May. Across 128 sessions, Opus 5 agents built and revised linked wiki pages; performance on 25 questions, answered from the wiki alone, rose from 13% to 48% of graded criteria. When Claude used both the wiki and a search tool, its aggregate score increased from 54% to 61% as the wiki grew, while search alone scored 53%.

A separate test used documents about a merger and the Harvey LAB benchmark, which asks Claude to generate legal briefs from source documents. One test asked 24 questions about the merger; more than half needed information from several documents, and many also needed judgment or math. On those items, Claude scored 87% with the wiki and 73% with search alone.