What Anthropic published
On November 6, 2026 Anthropic's engineering team published "The components of agent memory", written by Lance Martin. It describes memory as knowledge Claude gains from experience and keeps between sessions as context it can read later. Agents can organize that memory "by reflecting on logs to fix mistakes or produce new notes and representations. This process is often called dreaming."
The post does not announce a new Claude feature or a release date. It points to Anthropic cookbooks for trying dreaming and knowledge wikis yourself.
What Anthropic reports
On a corpus of about 350,000 internal files (roughly 420M tokens) from Stainless, a company Anthropic acquired, Opus 5 agents built a linked wiki over many sessions. Anthropic writes: "As we scaled dreaming from two sessions to 128, performance on the 25 questions grew from 13% to 48% of graded criteria." With both search and the wiki, Claude's aggregate score rose from 54% to 61%, compared with 53% with search alone.
On CL-Bench 1.0, a memory benchmark, Anthropic says "Claude Opus 5 was the first of these models to close the loop", meaning it turned wrong answers into tested rules, and that Opus 5.5 and Sonnet 5.5 did the same. In a 24-question merger test, on the items that needed combining several documents, judgment or math, Anthropic reports 87% with a wiki versus 73% with search alone; on one legal-brief task the wiki showed no clear gain. All of these are Anthropic's own evaluations, not independent tests.
Practical lessons in the post
Anthropic found that "it helps to tell Claude what memory is for, but not how to use it." When it told Opus 5 and Opus 4.7 to "record only failures", both lost much of the boost memory gave them. Memory loaded into every session must stay small: Claude Code's auto memory loads only the first 200 lines or 25 KB of MEMORY.md, which acts as an index.
What it means for you
If you use Claude Code or build agents, keep the always-loaded memory file short, describe the purpose of memory instead of rigid rules, and periodically review stored notes for errors. Do not expect a "dreaming" switch in Claude today: the post describes research and evaluation methods, and Anthropic has not announced availability or a timeline for users. Anthropic also notes that dreaming uses extra compute and that stored memories can drift out of date or be wrong from the start, so how often to run it and what it costs to maintain still need evaluation.
Go to the source
Sources & further reading
Based on the official documentation below. Features and access can change; check the provider before making a decision.
Was this page useful?
Help us make our coverage more useful. This is feedback on the page, not a rating of the tool.
Go a little further