Anthropic Explains How Agent Memory “Dreaming” Can Correct Notes
Anthropic’s November 6 engineering post by Lance Martin defines agent memory as retained readable context and describes background dreaming that reflects on logs to correct notes and build linked structures, based on internal evaluations.
Published
Last updated / Fact-checked on
Illustrative workspace, not Anthropic staff, a product demonstration or an evaluation setup. Photo: StartupStockPhotos / Pixabay, Pixabay Content License.
On November 6, 2026, Anthropic published an engineering post by Lance Martin titled “The components of agent memory.” The post defines agent memory as experience retained between sessions as readable notes, files or database entries. It distinguishes this from weight changes, single-session in-context learning and context compaction.
The post describes “dreaming” as idle background compute that reflects on logs to correct notes, build linked wikis, promote notes into injected context and target memory to a task. It credits the sleep-time compute idea to earlier Letta and UC Berkeley work. Dreaming is not framed as consciousness, literal sleep, online weight updates or a new universal product feature.
At a glance
| Aspect | Key point from the post |
|---|---|
| Memory | Retained readable context across sessions; excludes weights, in-context learning, compaction. |
| Dreaming | Background reflection on logs to correct notes, build wikis, promote content, target a task. |
| Pokémon (one run each, ~10k steps) | Sonnet 3.7: 413 files in 225 folders, one badge. Fable 5: nine files in two folders plus global.md, four badges. |
| Stainless (vendor) | ~350k files / ~420M tokens, 25 questions. Graded criteria 13% to 48% (2 to 128 sessions). Search 53%; search + wiki 54–61%; ~20k tokens saved per question (search + wiki). |
Dreaming as source verification and representation maintenance
Anthropic describes using dreaming to check notes against original transcripts. Notes written on the basis of in-session salience can be wrong from the start or can drift. The post treats dreaming as a maintenance step rather than a form of awareness.
In one Pokémon anecdote, an incorrect note placed the Mt. Moon PokéCenter on the wrong side. The agent walked off a one-way ledge and could not progress. Dreaming compared the note to the transcript, corrected the location, and the agent then reached Cerulean City. Uncorrected or no-memory runs failed to get past the mountain. This is a task-specific example, not evidence of general reliability.
On an internal Stainless corpus of roughly 350,000 files and 420 million tokens, Opus 5 wiki-building agents ran across up to 128 sessions coordinated by a shared log. Graded criteria met rose from 13 percent to 48 percent as sessions scaled from two to 128. The source reports these as graded criteria, not percentages of questions answered correctly.
Distinctions among vendor-reported conditions
The post keeps experimental conditions separate. Search alone met 53 percent of graded criteria on the Stainless set. Search plus a growing wiki reached 54–61 percent and saved about 20,000 tokens per question. Wiki-only peaked at 48 percent. The 13-to-48 percent movement and the search-plus-wiki figures should not be collapsed into one improvement.
In a merger task of 24 questions, a wiki met 87 percent of criteria versus 73 percent for search on items needing synthesis, judgment or math. The gain is limited to those items. One LAB task mainly requiring single-document retrieval showed no clear gain. On another LAB task, the wiki grew larger than the source documents. In a contract task a shared table retained only 41 of 57 facts (72 percent) after agents overwrote one another. Revised instructions—one agent per contract page and verbatim quotes—allowed 57 of 57 criteria to be met. The post presents the improvement as an instruction and infrastructure change, not as proof that memory alone fixed accuracy.
On the CL-Bench 1.0 database task, each question started in a fresh Managed Agents session with a persistent memory directory. Prompting Sonnet 4.6 to read memory and test theories nearly doubled its score; exact endpoints are not stated in prose. A “record only failures” rule reduced much of the memory gain for Opus 4.7 and Opus 5. All results remain vendor-run.
Limits and open questions
DailyMech performed no independent reproduction. The private Stainless corpus and small question sets limit what these examples establish beyond their setups. Anthropic says maintaining memory requires compute, outputs from dreaming may not be immediately accessible, and longer instruction files reduce adherence. It also warns about incorrect or stale notes and accumulated hallucinations. The post says Claude Code loads only the first 200 lines or 25 KB of MEMORY.md, not its entire memory store. Transcript retention and privacy controls are questions for future reporting, not violations established here.
What to watch
Independent tests on public or larger datasets would clarify how far the reported effects extend beyond Anthropic’s internal setups. Total background compute required to keep notes current, and whether correction holds when environments change faster than in the described evaluations, remain open measurement questions. The post does not claim unrestricted production readiness.
How we reported this
This draft rests on a full reading of the primary Anthropic post by Lance Martin dated November 6, 2026. No hands-on testing or independent reproduction was done. This article relies on that primary account, not an independently reproduced evaluation. All figures and examples are attributed to Anthropic’s reporting.
Published and fact-checked November 8, 2026.
Sources
- Anthropic Engineering, “The components of agent memory,” written by Lance Martin, November 6, 2026: https://www.anthropic.com/research/components-of-agent-memory (publicly available; no clock time or timezone supplied in source)
DailyMech used AI tools to help research and draft this article; it was fact-checked by our editors.