Anthropic's 'Dreaming' Lets Claude Fix Its Own Memory Notes

In Anthropic's Pokémon Red test, Claude stood outside a Pokémon Center near Mt. Moon, believed it was somewhere else, and walked off the same one-way ledge three times. The cause wasn't the game. It was a single wrong line in its own memory notes.

The fix was Anthropic's dreaming, as the company calls it: described in an engineering post published Nov. 6, 2026 by Lance Martin. While the agent is idle, a separate pass rereads its notes against the original transcripts and rewrites what's wrong. It's a research write-up with benchmarks, not a product launch, and the numbers below are Anthropic's own.

Memory that isn't weights, context or compaction

Anthropic defines agent memory as knowledge Claude gains from experience and keeps across sessions. It leaves out model weights, in-context learning and compaction. What's left is files the agent writes and later reads.

Dreaming is the maintenance job on those files. Per the post, it reviews logs and transcripts offline, corrects errors, writes new notes, builds wikis and moves important details into the layer loaded in every session. The outlet Interesting Engineering covered it on Nov. 6, noting that Anthropic itself says the process has little to do with human dreams.

Two Pokémon runs, two ways to remember

Anthropic compared two models around step 10,000 of Pokémon Red. Claude Sonnet 3.7 left behind 413 files in 225 folders, about 977,930 characters, and earned 1 badge. The post says it "assumed everything might matter and wrote to memory indiscriminately."

Claude Fable 5 kept 10 files, about 43,406 characters, and earned 4 badges. One of its notes reads "If 3+ balls fail in a row, VARY the timing," which is a lesson, not a log. Our math: Fable 5's memory is about 4.4% the size of Sonnet 3.7's, or roughly 22 times smaller, with four times the badges. Anthropic adds that Opus 5.5 finished the game with a small notes folder, though it gives no figures for it.

The ledge that dreaming removed

The Mt. Moon run is the cleanest demo. A raw-notes run kept hitting the same ledge and never got through. After a dreaming pass compared the note with the transcript, found the wrong Pokémon Center location and corrected it, the agent reached Cerulean City at action 1,935. At one point, the agent in the failed run wrote, "I'm going to stop rather than keep looping."

A 350,000-file wiki, and what it beat

The harder test used a corpus of about 350,000 files, roughly 420 million tokens, from Stainless, a company the post says Anthropic acquired in May. Opus 5 agents built a wiki over up to 128 sessions and were graded on 25 questions.

The wiki alone climbed from 13% to 48% of graded criteria between 2 and 128 sessions. Plain search with no wiki scored 53%. So the wiki by itself never beat search. The real gain was combining both: 61% with the wiki after 128 sessions, which Anthropic says also saved about 20,000 tokens per question. The wiki grew from 103 pages and 766 links after two sessions to 222 pages and 12,893 links after 128.

On a 24-question merger set, the wiki scored 87% on items needing several documents, judgment or calculation, versus 73% for search alone. That's 14 points. Anthropic also reports a limit: on one of the legal briefs in the Harvey LAB benchmark the wiki did not beat search.

Why the newest models save more

On CL-Bench 1.0, a database task, a memory store lifted every model, but by very different amounts:

Sonnet 4.6 barely moved: +0.034, against +0.304 for Sonnet 5 (our subtraction). Anthropic says Opus 5 "was the first of these models to close the loop," meaning it read its memory and used it. Sonnet 5 checked its notes before answering 9 of 10 times, while Sonnet 4.6 did so less than half the time. A refined instruction pushed Sonnet 4.6 to 0.641, so part of the gap is prompting, not only model skill.

Telling Claude to "record only failures" backfired. The gain dropped by 0.234 for Opus 5, 0.150 for Opus 4.7 and 0.020 for Sonnet 4.6. Anthropic's advice: "tell Claude what memory is for, but not how to use it."

What to try before you build a dreaming job

The post doesn't release a dreaming recipe, trigger or schedule, so there's nothing to copy yet. What it does support is cheap to test. Give your agent a memory folder, state its purpose in one sentence, and skip rigid rules like "log only errors." Then, every so often, make a second pass that checks notes against transcripts, as in the Mt. Moon fix.

One more thing to check on your own setup: Anthropic notes Claude Code's auto memory loads only the first 200 lines or 25 KB of MEMORY.md. If your index is longer, the tail is invisible to the agent, and no amount of dreaming helps what never loads.

Source: Anthropic Engineering

anthropic claude agent memory dreaming ai agents

AI

About the author

AI Briefing Daily Editorial. The editorial desk of AI Briefing Daily, an independent publication covering artificial intelligence. Every story is researched against primary sources, checked by hand and reviewed by a person before publication.