Anthropic says the key to agent memory is knowing what to forget, not how much to store
Reported by The Decoder1 source ↓
Anthropic studied how Claude retains knowledge across sessions. Its tests found selective notes outperformed exhaustive archives, while reviewing past sessions helped correct faulty memories.
Key points
- In Anthropic’s Pokémon test, Claude Sonnet 3.7 made 413 files and earned one badge; Claude Fable 5 used a few files and earned four
- Anthropic says Dreaming checks stored notes against past session transcripts and can correct errors
- In a legal test, Claude scored 87% with a wiki and 73% with search alone
- The Pokémon comparison covered one playthrough per model
Details
Sonnet 3.7 created 413 files in 225 folders and finished with one badge, while Fable 5 kept a few files and earned four. Anthropic said Fable 5 recorded plans and tested strategies rather than information already visible on screen; the comparison covered only one playthrough per model.
Anthropic also described Dreaming, which reviews earlier session transcripts and checks them against stored notes. In the Pokémon test, Claude had recorded the wrong side of a mountain for a landmark and repeatedly followed that faulty note; after reviewing the transcript, it corrected the error and found the route. Dreaming can also reorganize knowledge bases, though Anthropic says the process uses extra compute and takes place after the task.
In a separate legal test reported by Anthropic, Claude scored 87% with a wiki built from knowledge across documents, compared with 73% using search alone. The wiki helped most when answers were spread across multiple documents and performed roughly on par with search for simple lookups. Anthropic says notes and wikis require upkeep, and loading too many memories into every session can consume context and carry errors forward.
Written from coverage as of
Images
- The Decoder
- The Decoder
- The Decoder
- The Decoder
Stocks in this story
Change since open: 20 Nov to 21 Nov 2026
Last refresh