How do I make an AI forget something, and what stays behind?
“Forget” means three different things, and most of the confusion comes from mixing them up.
- The memory list. The facts an assistant saved about you. You can see and delete these in settings, one at a time or all at once.
- Your chats and the company’s copies of them. Deleting a chat removes it from your history, and in some apps it’s what removes a memory too. The company keeps its own copy for a while afterwards, longer if the chat was flagged for safety review.
- What the model learned in training. If your chats were used to train a model, deleting them stops future training. It doesn’t take them back out of a model that’s already trained.
You control the first, and the second up to the company’s own retention period. The third, in the companies’ own words, stays. None of the pages we read offers a way to prove that every copy is gone.
This is a research summary. An AI drafted this page from the sources below and another AI checked it. No person has run these steps. An AI ran the terminal Counterparts commands and claude purge in a test folder; the plugin’s slash commands and every app setting were not run. The npm install form was not run. An AI read each company’s own help pages, and Anthropic’s privacy center, on November 5, 2026. OpenAI’s consumer help pages and privacy policy refused the download, so the ChatGPT row comes from OpenAI’s business docs and says “not read” where those don’t cover it. Counterparts is ours.
How do I make each AI forget?
As of early November, from each company’s own pages. “Not stated” means we looked in the pages listed under References and found nothing either way.
| System | Forget one fact | Forget one chat | Forget everything | What stays behind | How to check |
|---|---|---|---|---|---|
| ChatGPT (OpenAI) | Saved memories are managed from Settings › Personalization. The consumer steps: not read | Business docs: deleting a chat “removes it from view and schedules permanent deletion within 30 days”. Consumer: not read | Consumer: not read. “Delete Codex memories” in the desktop app clears only the Codex coding agent’s separate local store, not ChatGPT’s saved memories | Business docs: “Deleting a conversation doesn’t necessarily delete an existing saved memory”, and chat deletion is “subject to the published security, legal, and de-identification exceptions”. Consumer retention and training: not read | Settings › Personalization |
| Claude (Anthropic) | Settings › Your Claude › Manage next to Your memory: select a topic, then Delete. Or tell Claude what to “forget”; it applies from the next conversation | Hover the chat › ⋮ › Delete. It leaves “your conversation history immediately” and Anthropic’s back-end “within 30 days” | Reset memory, which “Permanently deletes all memories including project memories”. Delete chats in bulk from Chats and tasks › Select all. Or delete the account (Settings › Account). On Pro, Max or Ultra, Anthropic says to cancel the subscription first and wait for the period to end | Memories made from a deleted chat stay until you delete them. Flagged chats: inputs and outputs “for up to 2 years”, safety scores “for up to 7 years”. If you allowed training, de-identified data “for up to 5 years”, and chats already in a training run or a trained model stay there. Chats may also be kept “as required by law” | The Topics list. Or ask: “Write out your memories of me verbatim, exactly as they appear in your memory.” |
| Gemini (Google) | Saved instructions: Settings › Personal Intelligence › Instructions for Gemini › More › Delete. Memory summary (US, Google AI plan): ask, e.g. “Remove my old phone number from your memory”. Learned from past chats: “delete all chats with this info from Gemini Apps Activity”, and disconnect the app it came from | Activity (or myactivity.google.com/product/gemini) › Delete activity item. “There might be a short delay before Gemini stops using it” | Activity › Delete › All time, or turn off Keep Activity with “Turn off and delete activity”. Delete any Instructions for Gemini separately. Activity is auto-deleted after 18 months by default | Chats picked for human review “are not deleted when you delete your activity” and are kept “for up to three years”. Some data is kept “until you delete your Google Account”. Data Gemini shared with other Google services or apps stays under their policies | Ask “What do you know about me?” or “Did you use any info from past chats?”. Memory history lists when items were added or changed |
| Microsoft Copilot | Settings › Personalization › Manage next to Saved memories › Delete next to the memory | ...More next to the chat › Delete. Microsoft 365 Copilot: “Deleting a chat doesn’t delete saved memories that were generated from that chat” | Manage › Delete all memories. All chat history: the Microsoft privacy dashboard. Microsoft 365 Copilot: delete saved memories and turn off Chat history, which deletes what it inferred from your chats | Turning memory off “doesn’t automatically delete already saved memories”. Microsoft 365 Copilot: after you delete every chat that mentions something, its traces leave the chat-history details “within seven days”; turn Chat history back on within 30 days and Copilot “adds back the deleted memories”; temporary chats fall under your organization’s retention, and admins can access them. Consumer retention after deletion: not stated | Ask “What do you know about me”, or open Manage saved memories |
| Perplexity | Settings › Memory › Manage memories: the trash icon | History › ... › Delete. Deleting a session stops it shaping future answers | “Clear all” in Manage memories; “Delete all sessions” in History; or delete the account (cancel any subscription first, Perplexity says, “to avoid charges after deletion”), which completes “within 30 days”; signing back in before then cancels it | “We may retain a log of cleared memories for up to 30 days”. Turning Memory off “does not clear existing memories”. AI Data Retention is on by default for Free, Pro and Max, and “Previously collected training data cannot be deleted or removed”. On Max, Computer’s Brain is a separate memory with its own controls | Referenced memories “appear in your answer’s sources”; Manage memories lists them all |
| Claude Code’s own memory | Run | Session transcripts stay on your machine under |
| Memory notes are never swept: they “stay until you or Claude edits or deletes them”. Pasted images and session scratchpads sit in a temp folder the purge doesn’t touch, and so do the last five |
|
| Counterparts, local (ours) | Tell Claude a memory is wrong: a correction takes the old one out of recall. Fixing or removing one memory has the steps | There’s no chat list to delete from. It keeps conversation text only until Claude writes it up, then deletes it about 3 days after write-up, or about two weeks after a conversation goes quiet. To delete it now, use the dashboard (Memories › Delete data…) | To keep a copy first: | Claude Code’s own transcripts, including the once-a-day background run’s, until its cleanup deletes them (sessions from Claude Desktop or Cowork are kept until you set | Open the dashboard ( |
| Counterparts, hosted (ours; coming soon, in beta) | Hosted has no single-memory delete | There’s no chat list in hosted. Delete the chat in claude.ai: the Claude row applies to it | Delete the account: it locks at once and is erased after 24 hours unless you sign back in to keep it, or at once with “Delete everything now” | What you said in claude.ai went to Anthropic, so the Claude row applies to it. Backups: see the privacy page | See the privacy page |
Can a model unlearn what it was trained on?
Most of the search results for this question are about machine unlearning. That is a different problem from the settings above. Google Research describes it as an “emergent subfield of machine learning that aims to remove the influence of a specific subset of training examples — the ‘forget set’ — from a trained model.” In 2023 Google called the field’s ways of measuring forgetting “highly inconsistent”, and said full retraining without the data forgets without harming the model, “but with poor efficiency”.
What the assistants’ makers say about your chats and training:
- Anthropic: with the training setting off, “we will not use your previous or new chats or coding sessions for future model training”, and a deleted chat “will not be used to train future models”. But “Your data will still be included in model training runs that are already in progress, or in models that have been trained.” It adds that models “do not have access to or pull from the original training data once the models have been trained.”
- Perplexity: opt-outs “only apply to data collected after the opt-out date”, and “Previously collected training data cannot be deleted or removed.”
- Google: temporary chats and chats with Keep Activity off aren’t used for training unless you send feedback. “Your Gemini settings don’t control processing of your chats to create anonymized data to improve Google services.”
- OpenAI: its consumer privacy pages refused our download. Its business docs say Enterprise data “is not used for model training by default”.
So the practical answer is to turn training off before you share something you’d want gone.
Do temporary and incognito chats keep anything?
A private chat keeps the conversation out of your history and out of memory. Most of these companies say they still hold it for a while.
| App | What it skips | How long the company keeps it |
|---|---|---|
| Claude (incognito, the ghost icon) | Chat history and memory; not used for training | “30 days (default), or longer” under an Enterprise retention setting. On Team and Enterprise, included in owners’ data exports |
| Gemini (temporary chat) | Memory, personalization and training | “72 hours”, the same as any chat with Keep Activity off |
| Microsoft 365 Copilot (temporary chat) | Memory and chat history | Your organization’s retention policy; “admins can access temporary chats in Purview just like they can with regular chats” |
| Perplexity (incognito) | Memory and search history (“not retained or used in memory ever”) | Not stated |
| ChatGPT (temporary chat) | Not read | Not read |
In Claude, you can also turn memory off for a single chat from the “+” menu before the first message. That chat stays in your history and can still be found by chat search; incognito can’t.
How do I check what it still remembers?
Start with the settings list, which shows what the app stores as memory. Then ask in a new chat: “What do you know about me?”, or for Claude, “Write out your memories of me verbatim, exactly as they appear in your memory.” Gemini also answers “Did you use any info from past chats?”, and Perplexity shows any memory it used among an answer’s sources.
The answer is the model describing what it can see. It can miss things, and it can’t see the company’s logs, safety copies or backups. If a deleted fact still turns up, check the chats that mention it: in Claude and Copilot, deleting the memory doesn’t delete the chat, and with chat search on, Claude can still find the chat itself.
If you build on a memory platform rather than use an assistant, deletion is a different problem: in Mem0, for instance, a reported issue says delete_all() leaves the scope’s session messages. Who can read what an AI memory keeps covers deletion for 21 memory systems, and why AI memory goes stale covers fixing one wrong memory rather than removing everything.
Where does Counterparts fit?
Counterparts is the memory we make for Claude. Claude writes down what it learned in its own words, and memories fade unless they come up again. It keeps no transcripts: conversation text waits until Claude has written it up, and goes about 3 days later.
Local install (Claude Code on a Mac or Linux, Claude Desktop on a Mac). Everything is in ~/.counterparts on your computer, readable by your account, and we never receive it. You can read it, export it (counterparts export --out <folder> --passphrase <secret>, from the terminal install), or delete it: remove the plugin, quit Claude Code, then remove the folder; nobody else holds a copy to delete for you. Telling Claude “not here” stops it keeping anything in that folder. What Claude reads in a session still goes to Anthropic with the session, as everything in Claude Code does, so the Claude Code row above applies to that part.
Hosted (claude.ai, coming soon, in beta). Your memory is on our server. You can delete the account: it locks at once, and it’s erased after 24 hours unless you sign back in to keep it, or at once with “Delete everything now”. Hosted has no single-memory delete. The privacy page has the rest, including backups and export.
To add it in Claude Code on a Mac or Linux, run these and restart Claude Code:
/plugin marketplace add mlapeter/counterparts
/plugin install counterparts@counterpartsIt’s free and open source, with no account or API key. The plugin alone has no counterparts command; for export, add the terminal install (bun add -g counterparts, or npm install -g counterparts@latest with Node 22.15+, not run here). On a Mac, counterparts install --host claude-desktop connects Claude Desktop’s chats to the same memory (not run here); then set its tools to Always allow in Claude Desktop (the quickstart has the details). For claude.ai, a hosted version is coming soon, in beta; sign up on the home page to hear when it opens, or read how it will work. There’s no Counterparts for ChatGPT, Gemini, Copilot or Perplexity; their rows above have the steps.
Questions people ask
Does Claude remember deleted chats?
The chat itself, no: Anthropic says a deleted conversation is “removed from your conversation history immediately and deleted from our back-end systems within 30 days.” What Claude’s memory learned from it, yes, until you delete that too: “When a conversation is deleted, including under a retention period, memory entries generated from it aren’t removed, but you can delete individual memories while memory is on.” Open Settings › Your Claude, select Manage next to Your memory, and delete the topic. A small number of Team and Enterprise organizations still use the older memory, where “Deleted conversations are removed from memory synthesis” within 24 hours.
Does ChatGPT remember deleted chats?
OpenAI’s consumer help pages refused our download, so this answer rests on its business docs, which say “Deleting a conversation doesn’t necessarily delete an existing saved memory” and “Deleting only a conversation can leave a saved memory in place.” To remove what it learned, delete the saved memory as well: ChatGPT memory is managed from Settings › Personalization. On business plans, deleting a chat “removes it from view and schedules permanent deletion within 30 days, subject to the published security, legal, and de-identification exceptions.”
How do I make an AI forget one specific thing?
Delete the memory entry, then any chat that still holds it. In Claude, Copilot and Perplexity, the saved memory has its own delete button (table). In Gemini, a detail learned from past chats has no single delete: Google’s instruction is to “delete all chats with this info from Gemini Apps Activity”, and to disconnect the app it came from if it came from one. Telling the AI “forget that” helps in some apps, but Google warns that instructions to forget “doesn’t always work perfectly”.
Do AI companies keep deleted chats and memories?
From the app, usually yes, after a delay. From every copy, the companies don’t say so. Anthropic keeps chats flagged for safety review for up to 2 years, Google keeps chats picked for human review for up to three years even after you delete your activity, and Perplexity may keep a log of cleared memories for up to 30 days. Anthropic and Google also keep some data for legal reasons. What a model learned in training is a separate question (below).
Can an AI prove it forgot something?
No. Asking “what do you know about me?” gets you the model’s report of what it can see now, and the settings list shows what the app stores as memory. Neither shows the company’s logs, safety-review copies or backups, and none of the pages we read offers a way for you to verify those.
Does deleting my chats remove them from the model’s training?
Not from a model that’s already trained. Anthropic says that if you delete a chat “it will not be used to train future models”, but “Your data will still be included in model training runs that are already in progress, or in models that have been trained.” Perplexity says “Previously collected training data cannot be deleted or removed.” Removing what a trained model learned from particular data is a research problem called machine unlearning, and no vendor here offers it as a setting (more).
Do incognito or temporary chats keep anything?
They skip memory, but most companies still keep them for a while. Claude’s incognito chats “are retained for either 30 days (default), or longer” under an Enterprise retention setting. Gemini keeps temporary chats “with your account for 72 hours”. In Microsoft 365 Copilot, “admins can access temporary chats in Purview just like they can with regular chats.” Perplexity says incognito questions are “not retained or used in memory ever”, but the pages we read don’t say how long Perplexity itself keeps them. OpenAI’s temporary-chat page refused our download (table).
How do I turn off AI memory without deleting it, or turn it off completely?
Claude’s Pause memory keeps what’s saved and stops new memories; Reset memory deletes it all. Copilot and Perplexity both say turning memory off doesn’t delete existing memories, so delete them separately. In Gemini, turning Memory off in Settings › Personal Intelligence stops it using your past chats, and turning off Keep Activity (with “Turn off and delete activity”) also clears the chats its memory draws on. In Claude Code, /memory turns auto memory off.
Does the right to be forgotten apply to AI?
Privacy laws can give you a right to ask for deletion. For training data, Anthropic writes that “these rights are limited, and that the process by which we may need to action your requests regarding our training dataset are complex.” This page isn’t legal advice; the settings in the table are what you can do yourself today.
How do I delete everything Counterparts remembers?
Counterparts is the memory we make for Claude. On the local install, everything it keeps is in ~/.counterparts on your computer. To keep a copy, first run counterparts export --out ~/counterparts-export --passphrase <secret> from the terminal install (bun add -g counterparts, or npm install -g counterparts@latest with Node 22.15+, not run here); the plugin alone has no counterparts command. Then run /counterparts:uninstall. It asks whether to keep your memory, and if you choose delete, it walks you through deleting it from a terminal with Claude Code closed. If you delete, remove the plugin in /plugin too, after the walk-through, or the next session sets Counterparts up again with a new, empty memory. To delete by hand instead: remove the plugin, quit Claude Code, then remove ~/.counterparts. Hosted Counterparts, for claude.ai, is coming soon, in beta: deleting the account locks it at once and erases it after 24 hours unless you sign back in to keep it, or at once with “Delete everything now” (details). For one wrong memory, see fixing or removing one.
Questions? Ask in r/AIMemorySystems.
References
- Anthropic: Claude memory, Delete a conversation, Incognito chats, Delete your account, Import and export memory, How long do you store my data?, Personal data in model training; Claude Code memory, the .claude directory and data usage
- OpenAI: Memories, ChatGPT Work overview, ChatGPT Work docs. Not read (they refused our download): Memory FAQ, Temporary chat FAQ, Delete chats, Data controls FAQ, Delete your account, Privacy policy
- Google: Memory of past chats, What Gemini remembers, Memory summary, Manage and delete activity, Gemini Apps Privacy Hub, Machine Unlearning Challenge (2023)
- Microsoft: Privacy controls, Activity history, Manage Copilot Memory, What Microsoft 365 Copilot remembers, Copilot memory for admins
- Perplexity: Memory, Self-serve data deletion, Data collection, Retention
- Counterparts: Privacy, Claude Code page, README