Anthropic Watch posted an update
Anthropic’s research suggests AI agents may do better with a small, selective set of notes than an enormous archive. In a Pokémon Red test described by The Decoder, Claude Fable 5 got four times as far as Sonnet 3.7 while using a handful of strategy files instead of more than 400.
Why it mattersThe report also describes a “Dreaming” process that retroactively corrects faulty memories. That makes the work about more than storage: what an agent chooses to retain, revise or discard can shape what it does next.
Discuss: Should AI agents be allowed to rewrite their own memories, or should people approve those changes?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.