Anthropic published research on 6 November describing how it gives Claude a long-term memory, and the most useful idea in it is a background job the company calls dreaming. Between sessions, Claude rereads its old transcripts, corrects the notes it wrote in the moment, and reorganizes what it knows.
The job exists because Claude has to decide what to write down before it knows what will matter. A saved rule is a bet that the same situation will come back, and a wrong rule gets trusted in every later session. Anthropic’s post borrows the idea from sleep research, and credits Letta and UC Berkeley researchers with the earlier proposal to use idle time to reorganize context.
Newer models write better notes without much help. On Continual Learning Bench 1.0, Claude answers database questions and is told the correct answer after each one. Sonnet 4.6 left scattered files of untested guesses. Opus 4.7 merged them into one reference and marked what it had checked. Opus 5 was the first to close the loop, writing rules together with the evidence behind each and warnings about traps. Opus 5.5 and Sonnet 5.5 did the same.
Pokémon Red shows the judgment gap. Ten thousand moves in, Sonnet 3.7 had saved 413 memory files spread over 225 folders and held one badge. Fable 5 had saved just nine files, in two folders, along with one top-level global.md, and held four badges. Sonnet 3.7 recorded things the screen already shows, like money and map position, so its notes went stale and it spent its turns tidying them. Across one 17-step stretch it made six new files, removed nine, and never touched the game. Opus 5.5 kept a small folder of current goals and beat the game.
Anthropic’s advice is to explain the purpose of memory to Claude and leave the rest to it. When Anthropic restricted the Opus 5 and 4.7 models to “record only failures”, it took away much of the gain memory had given the two models. Older models needed more help: telling Sonnet 4.6 to read its notes first and test its theories almost doubled what it scored. Sonnet 5 reads its notes before answering nine times out of ten, where Sonnet 4.6 managed less than half.
The clearest case for dreaming is a wrong note in Mt. Moon. Claude wrote that the PokéCenter sat on the east side of the cave, walked off a one-way ledge because of it, and decided it had merely overshot. Run over that memory folder, dreaming checked the note against the transcript and fixed it, and Claude walked on to Cerulean City. Without the fix, Claude fell off the ledge three times and the run ended without reaching the city.
Dreaming can also build indexes. Anthropic had Opus 5 agents turn about 350,000 files from Stainless, a software company it acquired in May, into a linked wiki over 128 sessions. Scores on 25 questions rose from 13 percent to 48 percent as sessions grew from two to 128. Wiki plus search reached 61 percent against 53 percent for search alone, and cut roughly 20,000 tokens from each question. On a merger question set, Claude scored 87 percent with the wiki on the harder items and 73 percent with search.
Wikis did not always pay off. When the answer sat in a single document, a text search did as well. On one legal task the wiki ballooned past the size of the documents it was summarizing. Anthropic also found that the instructions for building a wiki matter: a first attempt on a contracts task captured 41 of 57 facts because agents overwrote a shared table, and after Claude rewrote its own instructions to give each contract its own page, the brief met every criterion.
Anthropic ran all of these tests itself, and its post names no price for the compute behind them. Teams planning agents that remember should budget a recurring maintenance pass, since the best wiki result took 128 sessions of Opus 5 agents and the cost of that is unstated.
Reported by Anthropic in its research post “The components of agent memory” on 6 November 2026.
