Anthropic says the key to agent memory is knowing what to forget, not how much to store

Anthropic research shows that AI agents perform better by selectively storing key information rather than archiving everything, with Claude Fable 5 achieving four times better performance than Sonnet 3.7 on a Pokémon task using fewer stored files.
Why it matters
Understanding how to design agent memory systems efficiently could make AI assistants more capable and practical for long-running tasks.


