rotative
the decoder · 2026年11月16日

Anthropic 發現 AI 智能體記憶的關鍵在於選擇性遺忘

anthropic 測試了 claude 跨會話的記憶,發現選擇性記筆記優於詳盡的存檔,並且一個名為“dreaming”的過程可以糾正錯誤的筆記。這提高了智能體的性能。

read in English · 简体中文

anthropic tested claude's ability to store and retrieve information across sessions. claude wrote notes to a folder, but exhaustive note-taking overwhelmed the model with outdated data. selective note-taking, focusing on strategies and rules, worked better. anthropic also developed 'dreaming', a process where claude reviews earlier session transcripts during idle periods to correct faulty notes and build structured knowledge collections. this approach improved claude's performance in tasks like playing pokémon.

read at source ↗

← rotative