Anthropic describes Claude's memory, and the step it calls dreaming
On 6 November 2026 Anthropic published an engineering post on agent memory. Dreaming is the offline pass that reorganizes logs between sessions. It is not a weight update.
Anthropic published an engineering post on 6 November 2026 about how Claude keeps memory between sessions. The @AnthropicAI post that day points to the page “The components of agent memory.” The claim is narrow. Claude stores what it learns from work, and a separate pass, which Anthropic calls dreaming, reorganizes that store when the agent is not on the task.
Memory, on that page, is not a change to the model’s weights. It is not what Claude picks up inside one conversation. It is not the summary that replaces older turns so a session can continue. It is knowledge from experience, saved so a later session can read it.
The page draws four layers. At the bottom, the transcript, a searchable log. Above that, notes Claude writes. Then representations such as wikis. At the top, context loaded at the start of every session. Dreaming is the pass that moves material up from transcripts, and between the upper layers. The page says it can correct a mistake, write a note, build a new representation, or decide what is loaded before the next session starts.
What the page measured
One chart on the page follows a run toward the city Cerulean, over 3,000 cumulative actions. The run with memory plus dreaming reaches Cerulean at action 1,935. The run with raw memory falls off a ledge and backtracks to Pewter. The run with no memory never reaches Cerulean. The caption says dreaming corrects the memory.
Another test uses an internal corpus from Stainless, a software company Anthropic says it acquired in May. About 350,000 files, roughly 420 million tokens. The evaluation set is 25 questions. Over 128 sessions, Opus 5 agents explored the files and wrote or corrected linked wiki pages. On the chart, wiki plus search rises from 54% to 61% of graded criteria. Wiki alone rises from 13% to 48%. Search with no wiki stays flat at 53%. One example wiki goes from 103 pages and 766 links after 2 sessions to 222 pages and 12,893 links after 128 sessions.
On the Harvey LAB benchmark, the page describes a corpus of 10 documents and 115 KB of text. The wiki built from it has 68 pages, 255 links, and 203 KB, which the page puts at 1.77 times the corpus. When several agents wrote to the same shared page, 41 of 57 graded facts survived. With one page per contract, and the terms copied word for word, 57 of 57 survived. A brief graded from that improved wiki could meet 100% of the criteria, up from 72%.
The page also describes Claude Code’s auto memory. Claude does not load all of it into every session. It loads the first 200 lines, or 25 KB, of MEMORY.md. That header is an index.
The page says Opus 5.5 and Sonnet 5.5, like Opus 5, used memory to store validated rules on a Continual Learning Bench task. With memory, both scored about the same as Opus 5. Their baseline without memory was slightly higher. The extractor did not return the exact score. It is not in this draft.
What this desk did not open
This desk did not open Continual Learning Bench, the Harvey LAB repository, or the dreaming docs in the Claude console. The page does not say what a dreaming pass costs, or how often an ordinary Claude.ai session runs one. The post is an engineering note. It is not a product switch.