Anthropic Explains Agent Memory and Dreaming: Memory Lifted CL-Bench Scores for All Six Models Shown and Dreaming Corrected a Wrong Note and Built Wikis, but the Evidence Is Anthropic's Own and Dreams Are a Research Preview Without ZDR or HIPAA Eligibility

On November 6, 2026, Anthropic's engineering blog published The components of agent memory, written by Lance Martin, about how Claude agents keep knowledge between sessions. Anthropic defines agent memory as what Claude learns from experience and saves as context to read later. It leaves out changes to model weights, learning inside a single conversation, and compaction. The post describes four layers, from transcripts to notes to representations such as wikis to context loaded into every session. It adds a background process it calls dreaming, which reviews transcripts to fix and reorganize those layers. All results come from Anthropic's own tests.

The first test used CL-Bench 1.0, a benchmark where an agent answers a series of database questions and is told the right answer after each one. Claude started every question in a fresh session, with only a memory directory carried over, and the post compares that with a stateless baseline. Mean reward rose for all six models shown: Opus 5.5 from 0.655 to 0.838, Sonnet 5.5 from 0.644 to 0.833, Opus 5 from 0.564 to 0.828, Opus 4.7 from 0.383 to 0.700, Sonnet 5 from 0.306 to 0.610 and Sonnet 4.6 from 0.330 to 0.364. The post also describes how the habit improved. Sonnet 4.6 wrote scattered files of untested guesses. Opus 4.7 marked which facts it had checked. Opus 5 was the first to close the loop, recording a rule together with how it knew the rule was true, plus warnings about traps.

The second test looked at judgment about what to save. In Pokémon Red, the post compares one run each. About 10,000 steps in, Claude Sonnet 3.7 had written 413 memory files in 225 folders and had one badge. Claude Fable 5 had nine files in two folders, plus a top-level global.md, and four badges. The post's explanation is that Sonnet 3.7 recorded things the screen already shows, such as money and position, which went stale and cost time to search, while Fable 5 kept plans and tested rules that exist only if the model writes them down. Opus 5.5 also kept a small folder and beat the game. Two single runs are an illustration, not a statistical result.

On steering, Anthropic says it helps to tell Claude what memory is for but not how to use it. Simply telling Claude that a memory folder existed was enough for Opus 4.7 and Opus 5. Telling them to record only failures cost both much of the benefit that memory gave them. Older models gained from stronger prompts: telling Sonnet 4.6 to read its memory before answering and to test its theories nearly doubled its score, while Sonnet 5 read memory first in 9 of 10 cases without being told, against fewer than half for Sonnet 4.6.

Dreaming is the main new idea. Because Claude writes notes while it works, a wrong note can be trusted in a later session. In one Pokémon example, Claude wrongly noted that a building was on the east side of Mt. Moon, concluded the cave was behind it, and fell off a one-way ledge back onto its earlier route. After a dreaming pass checked the note against the session transcript and corrected it, Claude re-entered the cave and reached Cerulean City. In the run with the uncorrected note, Claude dropped back onto the route three times and never got there. The post also shows dreaming building wikis. On about 350,000 internal files, roughly 420 million tokens, from Stainless, a company Anthropic acquired in May, performance on 25 questions grew from 13% to 48% of graded criteria as Opus 5 agents built a wiki over 2 to 128 sessions. A wiki combined with search scored 54% to 61% as the wiki grew, against 53% for search alone, and saved about 20,000 tokens per question. On 24 merger questions, items needing several documents scored 87% with the wiki and 73% with search. On some legal tasks the wiki showed no clear gain, and on one it grew larger than the documents.

Two more results are about instructions. A wiki where several agents wrote to one shared table kept only 41 of 57 graded contract facts, or 72%. After Claude rewrote the build instructions in a reflection loop, giving each contract its own page quoted word for word, all 57 facts survived. A second dreaming pass also repointed nine links from a missing escrow-agreement page to an existing buyer-form-escrow-agreement page. The post warns that injected context must stay small, because every note costs tokens in every session and a wrong note travels everywhere. Claude Code's auto memory loads only the first 200 lines or 25 KB of its MEMORY.md index.

The Managed Agents documentation adds the product facts. A dream is an asynchronous job that reads an existing memory store and 1 to 100 past sessions and produces a new, reorganized store, without changing the input. It usually takes minutes to a few hours. The status is research preview, access is by request, it needs the beta headers managed-agents-2026-04-01 and dreaming-2026-04-21, and it is not eligible for zero data retention or HIPAA. The docs say instructions steer the synthesis, but imperative edits to specific lines generally produce no change.

Our reading: this is a useful map of design choices, a design and experiment report. The evidence is Anthropic's, the Pokémon comparison is two single runs, the wiki results come from one company corpus and a few legal tasks, and dreaming costs compute. Teams that build agents can borrow the ideas: keep the injected layer small, let strong models judge what to save, and check notes against transcripts. Deployments requiring Zero Data Retention or HIPAA eligibility cannot rely on the current research-preview dreams for those qualifications. We did not run any of these tests.