Anthropic发布Claude记忆管理文及清理提示词

Anthropic just published an engineering post on how Claude slims down and corrects its own memory. You can ask Claude Code to use it to review your project's auto memory top to bottom and clean it out.

Sonnet 3.7 played Pokémon and wrote everything down to memory. It piled up 𝟰𝟭𝟯 𝗳𝗶𝗹𝗲𝘀 and earned just 1 badge. Fable 5 wrote 𝗼𝗻𝗹𝘆 𝟭𝟬 𝗳𝗶𝗹𝗲𝘀 and earned 4. That's an example from the article.

Claude Code's auto memory is on by default. The longer you use it on a project, 𝘁𝗵𝗲 𝗱𝘂𝗺𝗯𝗲𝗿 𝗶𝘁 𝗳𝗲𝗲𝗹𝘀. That's because its memory is full of notes that were wrong or went stale when the code changed. It trusts that junk every time before it writes code.

Send this prompt and the link below to Claude Code 👇

"Read this article, then review this project's auto memory (MEMORY.md and the other files in the memory folder) against the current code and git history. Pick out anything you can read straight from the code, outdated notes that no longer match the code, guesses that were never verified, play-by-play logs with no conclusion, and entries that repeat or contradict each other.

For each one, quote the original, explain how keeping it would mislead you, and suggest the smallest change. List any memories to delete and any changes involving safety or permissions separately for me to decide. Just show me the recommendations. Don't change any files yet."