Agent Memory: Anthropic Details How “Dreaming” Helps Claude Correct and Organize Its Notes
In an engineering post published on November 6, Anthropic detailed its agent memory experiments: notes taken in-session, followed by off-session “dreaming” passes to correct errors, build wikis, and determine what loads at each startup.
Anthropic published a lengthy engineering post on November 6, 2026, titled “The components of agent memory”, exploring how its agents retain what they learn from one session to the next. The piece does not launch a product: it details experiments and a method—“dreaming”—that teams building agents can replicate.
Two Phases: Taking Notes, Then Consolidating
Anthropic defines “agent memory” as what Claude retains across sessions in a readable context format (such as notes in a folder), as opposed to model weights, in-context conversation learning, or compaction. The post separates this into two distinct stages:
- during the session, Claude takes notes on what it deems important;
- off-session, a dreaming pass reviews transcripts to correct those notes, write new ones, build a wiki, or decide what gets loaded by default next time.
The idea builds on sleep-time compute, proposed by Letta and UC Berkeley: using downtime between interactions to reorganize context.
What the Tests Show
On a CL-Bench 1.0 task (database queries with corrections provided after each answer), Anthropic highlights clear progress across generations: Sonnet 4.6 accumulates untested hypotheses, Opus 4.7 consolidates a reference document, and Opus 5 is the first to close the loop (a rule, how it knows it is true, and the pitfalls encountered).
In Pokémon Red, the difference lies in deciding what is worth remembering: around step 10,000, Sonnet 3.7 had created 413 files across 225 folders for just one badge, while Claude Fable 5 wrote nine files and a global.md for four badges. A specific example highlights the value of dreaming: an incorrect note regarding a building’s location caused Claude to walk in circles; once corrected against the transcript, the model cleared the obstacle.
On an internal Stainless corpus (roughly 350,000 files), the wiki compiled via dreaming pushed graded criteria scores on 25 questions from 13% to 48% as dreaming expanded from 2 to 128 sessions. Combined with a search tool, it reached 61%, compared to 53% for search alone, saving around 20,000 tokens per question. On a legal benchmark (Harvey LAB), the wiki helps when information is scattered across documents, but offers no advantage when searching within a single file suffices.
Recognized Limitations
The post notes that it is better to tell Claude why memory is used rather than how to use it: enforcing a rule to “only record failures” caused Opus 5 and Opus 4.7 to lose much of their performance gains. It also points out that the layer loaded at each session must remain small: Claude Code only reads the first 200 lines or 25 KB of MEMORY.md at startup. Finally, a wiki can expand until it outgrows the documents it summarizes, and dreaming incurs compute costs.
Our take
This post primarily serves as a design guide. For anyone running agents in production, three ideas are directly applicable: separating in-session note-taking from background consolidation, verifying notes against source logs rather than trusting them blindly, and keeping a short, preloaded index rather than a bloated instructions file. The metrics remain internal evaluations—sometimes based on single runs—and dreaming is not presented as an off-the-shelf feature. The key takeaway, relevant well beyond Claude: agent memory requires active maintenance; it cannot simply accumulate.
Sources
- Anthropic Engineering — The components of agent memory (November 6, 2026) Primary source
- Anthropic on X — announcement of the post (November 6, 2026) Primary source
A primary source comes directly from the company, lab or organization concerned. A secondary source reports the information without being its author: we flag it as such. Spotted an error? See our corrections policy.
