Describes a design for AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition → memory: capture salient events during sessions, then use offline 'dreaming' to fix mistakes, write notes and build new representations. Gives builders a concrete pattern and evalA repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.Full definition → approach for cross-session memory.
Key takeaways · AI-distilled
In Anthropic's design, Claude tags salient events or writes notes during a session, then an offline 'dreaming' pass checks notes against the original transcripts. In a Pokémon test, dreaming caught a wrong landmark note that had sent Claude over a one-way ledge.
On Continual Learning Bench 1.0, telling strong models only that a memory folder exists beat fixed rules: 'record only failures' cost Opus 5 and Opus 4.7 much of their memory gain. Prompting Sonnet 4.6 to read memory and test theories nearly doubled its score.
About 10,000 steps into Pokémon Red, Sonnet 3.7 had written 413 memory files in 225 folders and earned one badge; Fable 5 wrote nine files plus a global.md dashboard and earned four. Fable 5 stored plans, strategies and tested rules, not facts it could read off the screen.
Across 128 dreaming sessions, Opus 5 agents built a wiki from about 350,000 Stainless documents. Wiki-only scores rose from 13% to 48%, and search plus wiki hit 61% versus 53% for search alone. Wikis helped less when answers sat in one document.
On a contract-acquisition task, a wiki built by agents writing to one shared table met only 72% of brief criteria because agents overwrote each other. After Claude rewrote its build instructions to give each contract its own page quoting terms verbatim, the brief met 100%.
Terms in this piece · Glossary
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
eval — A repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.
