Anthropic finds compact memory can beat exhaustive logs for long-running agents

Anthropic has tested how AI agents should retain knowledge across long tasks, comparing raw logs, compact notes, structured wikis and an offline reflection process the researchers call “dreaming.” The results suggest that saving more context is not always better: agents benefit when they preserve validated rules and revisit earlier evidence to correct bad assumptions.

In a Pokémon Red experiment at roughly 10,000 steps, an older Claude agent created 413 files across 225 folders and earned one badge. A newer agent called Fable 5 kept nine files in two folders plus a global note and earned four badges. Anthropic says Claude Opus 5.5 later completed the game with a similarly small memory folder.

The team also gave agents access to an internal software corpus from Stainless containing about 350,000 files and 420 million tokens. Claude Opus 5 agents built a wiki over repeated sessions. Across 25 evaluation questions, the share of grading criteria satisfied rose from 13% after two reflection sessions to 48% after 128 sessions. In another example, reflection caught an incorrect navigation rule and replaced it before the agent repeated the same failure.

The practical lesson is that durable agent memory needs curation, testing and periodic revision—not just a larger context window. The study is still limited: it uses controlled game tasks and one internal corporate corpus, so it does not establish a universal memory design for production agents.