Anthropic's agent memory research: Fable 5 kept 9 files and got 4 badges, Sonnet 3.7 kept 413 and got 1
Anthropic published its findings on how Claude should keep memory across sessions, and most of them are about restraint.
What happened
In a Pokémon Red test, Claude Sonnet 3.7 saved nearly everything and ended up with 413 memory files across 225 folders and one badge at around 10,000 steps. Claude Fable 5 wrote nine files in two folders plus a global.md file, and earned four badges. Fable 5 skipped anything visible on screen and stored plans, strategies and rules it had tested. Anthropic compared one run per model, which The Decoder points out.
Anthropic found it helps to tell Claude what memory is for without dictating how to use it. A fixed rule to "record only failures" cost Opus 5 and Opus 4.7 much of their gain from memory, while older models benefited from explicit instructions.
The post also covers dreaming, where Claude reviews past session transcripts and corrects its notes. It fixed a wrong landmark note that had sent Claude back over the same ledge three times. On a wiki built from about 350,000 documents, combining the wiki with search raised scores from 54% to 61%, against 53% for search alone, and saved about 20,000 tokens per question. In a contract test, agents writing to one shared table kept only 41 of 57 graded facts because they overwrote each other. One agent per contract, quoting terms word for word, kept all 57.
My take
Agents that work in a CRM or helpdesk face the same choice as the Pokémon runs. The CRM already holds the deal stage and the last contact date, so copying that into memory just creates a second, staler copy. Memory should hold what the system of record doesn't: verified rules about how this client works, known exceptions and past mistakes.
The contract result is the one I'd act on first. Several agents writing to one shared note is a concurrency bug, and it silently drops facts. Give each record a single writer, and quote prices and terms exactly instead of summarizing them.
A scheduled review that checks notes against logs is also cheap to build, though Anthropic is clear that it costs compute and that these are its own tests, not independent results.
More posts
- Open OCR models now read 1,000 pages for under $0.10, and better serving saved more than switching modelsNov 20, 2026
- Gemini 4 Neon costs $3 in and $15 out per million tokens and tops Zapier's AutomationBench at 56.3%, though it uses more output tokens per taskNov 20, 2026
- Gemini API's Tool Retrieval loads tool schemas only when the model asks for them, so agents with hundreds of tools send fewer input tokensNov 20, 2026
