Dev News Daily ENDE

Anthropic describes how Claude's memory works and uses offline dreaming to correct wrong notes

Anthropic published a research post on 6 November about agent memory: knowledge Claude keeps between sessions as files it can read later, as distinct from training, in-context learning or compaction. The central idea borrows from sleep. Agents record their work as it happens, and a separate offline pass, which the company calls dreaming, reviews those logs to fix mistakes, write new notes and decide what is loaded into future sessions.

On the CL-Bench 1.0 database task, where the model learns the correct answer after each question and may write to a persistent memory folder, Anthropic says Claude Opus 5 was the first model to close the loop: its notes held a rule, the evidence for it and warnings about traps. Older models left untested guesses scattered across files.

A Pokémon Red run compared what models thought worth keeping. At about 10,000 steps Claude Sonnet 3.7 had written 413 files across 225 folders and won one badge; Claude Fable 5 kept nine files in two folders plus one index file and won four. In one case a note placed a landmark on the wrong side of a mountain; dreaming checked it against the session transcript, rewrote it, and the corrected run got through.

The post also reports that telling strong models only to record failures cut much of their gain from memory, and that wikis built by repeated dreaming over about 350,000 internal documents raised the share of graded criteria met from 13% to 48% across 128 sessions.

Anthropic describes how Claude's memory works and uses offline dreaming to correct wrong notes
Anthropic describes how Claude's memory works and uses offline dreaming to correct wrong notes — Dev News Daily

What it means

The useful lesson for anyone building agents is that a memory note can be wrong from the moment it is written and then be trusted in every later session. A pass that checks notes against the original record is what keeps memory from turning into a confident source of errors.