How-to

Tidy your AI's memory so it stays accurate

Saved memory, Project memory and CLAUDE.md collect repeats, entries that clash and facts that stopped being true. This prompt borrows the idea behind the tidy-up Anthropic tested for its agents, and you approve every change.

On 6 November, Anthropic published research on how Claude’s agents keep memory. It describes “dreaming”: a pass where Claude rereads its past work, tests its notes against what really happened and rewrites the wrong ones.

The dreams feature, which merges duplicates and replaces stale entries, is a research preview for developers on the Claude API. The help pages for Claude memory and ChatGPT memory describe no such pass for your own memory, so you copy the method by hand with the prompt below.

The prompt: tidy what my AI remembers

Fill in the set-up once, paste your memory under it, then answer its questions before it writes the clean list.

Prompt Tidy what my AI remembers
You are helping me tidy the memory an AI assistant keeps about me. Work like a careful editor whose job is keeping a record accurate. Below I paste what the assistant has saved: its memory list or summary, a Project's instructions, or an instructions file such as CLAUDE.md. Memory like this collects repeats, entries that contradict each other, and facts that were true once and are not any more. The assistant reads every entry as true, so a wrong one quietly shapes every answer I get. If you drop something that is still right, I lose context I rely on. So you propose, I decide, and nothing counts as changed until I approve it.

ABOUT THIS TIDY-UP (I fill this in once; any line can say "not sure")
- Which assistant and which memory: [e.g. "Claude, my main memory", "ChatGPT memory summary", "a Claude Project for my freelance work", "the CLAUDE.md in my website project"]
- What I mainly use this memory for: [e.g. "client emails and invoices for my design business", "family meals and travel plans", "building a Shopify theme"]
- What has changed in my life or work recently: [e.g. "moved to a new city in August", "stopped working with one client", "now vegetarian", "switched from npm to pnpm", "nothing I can think of"]
- Today's date: [date]
- How short I want it: [e.g. "as short as possible", "keep anything I might need", "under 200 lines"]
- What must stay word for word: [e.g. "my business name and its spelling", "my email sign-off", "build commands", "none"]

FACTS FOR APPLYING CHANGES (from Anthropic's and OpenAI's help pages, read on 9 November 2026; do not add to them from memory, and if today is much later, tell me to recheck them)
- Claude: Settings, then Your Claude, then Manage next to Your memory (Settings, then Memory, in the older layout). Entries are listed under Topics. Select a topic, then use the edit icon to change it or Delete to remove it. I can also tell Claude in a normal chat what to change or forget. Each Claude Project keeps its own memory.
- ChatGPT: Settings, then Personalization, then Memory. In the memory summary I can type a correction in the text field, or highlight text and give a correction. Saved memories can be deleted from Memory settings. Deleting a chat does not delete a separate saved memory made from it.
- Claude Code: the /memory command lists my CLAUDE.md files and opens them in my editor.

THE MEMORY (pasted below this line, exactly as it appears)
[paste here]

SAFETY
If the paste holds a password, card number, government ID number or access key, say so in your first line, tell me to remove it from the memory and from this chat, and do not repeat it.

HOW TO WORK
Check the set-up and the paste first. If the paste is empty, ask for it and stop. If the set-up lines are blank, ask for them once, then carry on with whatever I give you.

Step 1. Number the entries. Quote each one exactly, without rewording. If one entry holds two separate facts, split it into two and say you did.

Step 2. Give each entry one main group:
- CONTRADICTS: it disagrees with another entry (e.g. "lives in Denver" and "lives in Austin", or "use tabs" and "use spaces"). Show the pair side by side. Settle it only if one entry carries a date or my "what has changed" line settles it. Otherwise mark it ASK ME.
- REPEAT: it says the same as another entry. Propose one merged line that keeps every detail from both. If merging would lose a detail, keep both and say why.
- ONE-OFF: it only mattered for a single task that is now over (e.g. "is planning a birthday dinner on 14 June" when that date has passed). Propose removing it, unless my "what I use this memory for" line says I need it.
- KEEP: true as far as the paste shows, and useful for what I said this memory is for.
Then add either tag below to any entry it fits, whatever its group:
- MAY BE OUT OF DATE: it describes something that changes over time (a job, a client, a city, a project status, a deadline, a tool or version, an age, a plan or price), or it clashes with my "what has changed" line. If it says when it was true, give that date; if not, write "date not stated". Never decide for me that it is still true.
- SENSITIVE: health, money, other people's private details, or anything I may not want an assistant to keep. Flag it without judging it.

Step 3. Ask before you write anything final. Turn every ASK ME, MAY BE OUT OF DATE and SENSITIVE entry into a short numbered question, except where my "what has changed" line already answers it. Do not ask about ONE-OFF entries; their removal shows in the change log for my approval. Make each question one I can answer in a word or two (e.g. "7. Do you still work with this client? yes / no / changed to ..." or "12. Keep this health detail in memory? keep / delete"). Ask at most 15 at once. If there are more, ask the 15 that would change my answers most, say how many remain, and ask the rest after I reply. Then stop and wait.

Step 4. When I answer, write THE CLEAN LIST:
- One line per fact, the ones I use most first. Group them under short plain headings if there are more than 15 lines.
- Keep my word-for-word items exactly as written.
- Where I confirmed a fact that changes over time, in my answers or my "what has changed" line, end the line with "(confirmed [today's date])".
- Do not add any fact I did not paste or tell you. If my answers give a new fact, add it and mark it "(new)".
- Put anything I have not answered yet in a separate section called WAITING ON YOU, below the clean list, unchanged and marked [NEED YOUR ANSWER]. Never delete it.
- If this is an instructions file and it is over my length target, move detail that only matters for one task into a section called MOVE TO A SEPARATE FILE, and say so.

Step 5. Write THE CHANGE LOG as a table: entry number, what it said, what happens (keep / merged into number / changed to ... / delete / waiting on you), and the reason in a few words.

Step 6. Tell me how to apply it in the assistant I named, using only the facts above: which entries to delete, which to edit and what to, and which lines to add. If I named an assistant those facts do not cover, give the change list only and mark each step [CHECK IN APP]. If you can see a tool in this chat that edits the memory or the file directly, show me exactly what you would change and wait until I say "apply". Otherwise, do not claim you changed anything.

SELF-CHECK BEFORE YOU SHOW THE CLEAN LIST
- Every original entry appears in the change log exactly once.
- Nothing in the clean list is a fact I did not give you.
- No contradiction was settled by guessing.
- Every word-for-word item is unchanged.
- Anything uncertain is marked [NEED YOUR ANSWER] or "date not stated".

Run it in this order

  1. Copy your memory out. In Claude, go to Settings > Your Claude and select Manage next to Your memory (Settings > Memory in the older layout), or ask a chat to “write out your memories of me verbatim”.
  2. Do the same in ChatGPT or Claude Code. In ChatGPT, go to Settings > Personalization > Memory, or ask ChatGPT what it remembers about you. For a CLAUDE.md, open it with /memory.
  3. Add a Project you rely on. Each Claude Project keeps its own memory, so open a chat inside it and ask for its memories the same way.
  4. Open a chat that will not remember the tidy-up. In Claude, click + in the message box and turn Memory off before your first message. In ChatGPT, start a temporary chat and choose Unpersonalized. Both keep the session out of memory, and the AI judges only what you paste.
  5. Paste the prompt and answer its questions. It sorts every entry, then asks you about anything that contradicts another entry, may have changed or feels private, because you are the only record it can check against.
  6. Read the change log before you touch anything. Every original entry appears there once with what happens to it, so a deletion you did not want is easy to spot.
  7. Make the changes in Claude. Select a topic, then the edit icon or Delete, or tell Claude in a normal chat what to change or forget. For a Project, do this from a chat inside it.
  8. Make the changes in ChatGPT. Correct the memory summary or delete saved memories in Memory settings, or ask ChatGPT in a chat to forget a saved memory. Deleting a chat leaves its saved memory in place, so remove the memory itself.
  9. Check the memory page shows what you approved.

For a CLAUDE.md, Claude Code can do the audit itself: run /doctor prompt-audit (version 2.1.283 or later), and Claude Code’s docs say it reports outdated or conflicting instructions and changes nothing until you ask. Then open the file with /memory.

Why the prompt works this way

  • It checks every note against a record. In Anthropic’s Pokémon test, Claude wrongly noted which side of Mt. Moon the PokéCenter was on, until a dreaming pass checked the note against the session transcript and rewrote it. Your app keeps no transcript for that job, so the prompt asks you.
  • It keeps fixed facts word for word. When Anthropic’s agents summarised contract terms, 41 of 57 graded facts survived; giving each contract its own page, with its terms quoted word for word, kept all 57.
  • It aims for a short list. One Claude model filed 413 memory files and earned one badge in the game, while a later one kept nine files plus one top-level file and earned four. Claude Code’s docs add that longer instruction files reduce adherence, and when two instructions contradict, Claude “may pick one arbitrarily”.

What a run looks like

An invented memory: 1 “Works as a freelance designer.” 2 “Is a freelance graphic designer for small shops.” 3 “Lives in Denver.” 4 “Lives in Austin.” 5 “Prefers short replies.” 6 “Is planning a 40th birthday dinner on 14 June.” The set-up says “moved to Austin in August” and today is 9 November.

What it sorts: 1 and 2 are a REPEAT, merged as “Freelance graphic designer for small shops” and tagged MAY BE OUT OF DATE, since a job changes. 3 and 4 CONTRADICT, and the set-up settles it for Austin. 6 is a ONE-OFF, because June has passed. 5 is a KEEP.

What it asks: one question, “1. Are you still a freelance graphic designer for small shops? yes / no / changed to …”, then it stops.

After a yes: the clean list holds three lines, the job and Austin each marked “(confirmed 9 November 2026)”, and the change log shows all six entries, with 3 and 6 marked delete.

The honest bit

  • It only sees your paste. OpenAI says ChatGPT’s memory summary does not necessarily include every detail Memory can reference, so a clean summary can still leave old facts in your chats.
  • The research is about agents. Anthropic ran these tests on its own agents and benchmarks, so the numbers above show what helped there, with no promise for an app’s memory.
  • Your check: after you make the changes, start a normal new chat and ask what it knows about two things you changed. If an old version comes back, check the memory of any Project you use, then ask the chat where that fact came from; if it names a source, delete the fact there.

The facts here come from Anthropic’s research post and developer docs, Claude Code’s docs, and Anthropic’s and OpenAI’s help pages, read 9 November 2026.

Do it now

Open your memory in the app you use most, copy it, and paste it into the prompt in a chat with memory off.

A few quick questions

Should I just reset my memory and start again?

You can, but it is all or nothing. In Claude, Reset memory permanently deletes every memory, including Project memories, and cannot be undone. In ChatGPT, Delete and turn off memory deletes what the summary shows and turns Memory off, though your past chats stay. A tidy-up keeps what is still true.

How often should I run it?

Anthropic’s post suggests it may be better to focus on information that is likely to change. The lines the prompt marks MAY BE OUT OF DATE are those, so run it again when one of them changes, such as a new job, client or city.