AI memory: what it keeps, and a ten-minute audit
What ChatGPT, Claude and Gemini say they remember across chats, how to see it, and a ten-minute routine to list it, prune it and choose a memory setting for each kind of task.
- What this is
- What AI chat assistants save about you between conversations, how to see it, and a routine for clearing out what is wrong.
- Why it matters
- You avoid an assistant quietly reusing stale, wrong or sensitive details about you in answers where they do not belong.
- The short version
- Assistant memory can bring old details into unrelated chats and make the assistant agree with you more, so check what it holds.
- Ask it to list what it remembers, save the list, compare it with the memory settings page and delete what is wrong or sensitive.
- Keep memory on for casual use, use projects for long work, and use a temporary chat for health, money and other people's details.
On this page
Use it when you use one assistant with memory for more than one purpose, such as work, family and health in the same account, or when it brings up something you did not expect.
Not when you want the full privacy pass across training, retention and connectors (links that let an assistant reach your other apps). That is Privacy settings pass, which also lists where each vendor keeps its memory settings.
You need about ten minutes per assistant (our estimate). The audit and the table below are our suggestion and have not been tested.
Why memory needs checking
In a 2026 benchmark of 18 models, stored memories leaked into unrelated chats about half the time and reinforced users' biases almost always (median failure 53% and 97%) (PersistBench, peer-reviewed benchmark, ICML 2026). So memory can bring up the wrong thing, and it can make an assistant agree with you more. The benchmark tests models, not any one product's current memory.
Memory can also hold content you did not write. A 2024 attack planted instructions in ChatGPT's memory so that later chats leaked data. OpenAI patched the macOS app (The Hacker News, security report).
And it is partly invisible. OpenAI's memory FAQ says: "Unlike an explicit saved memory, information derived from chat history can change as ChatGPT updates what is most useful to remember." It says the memory summary "does not necessarily include every detail or source that Memory can reference", and that "Deleting the original chat does not automatically delete a separate saved memory" (OpenAI memory FAQ, vendor help). Deleting a chat is not the same as deleting what was learned from it.
The ten-minute audit
| # | Step | Time |
|---|---|---|
| 1 | Ask the assistant to list what it remembers, with the prompt below. | 2 min |
| 2 | Save the list as a file you own, such as memory-claude-2026-10-18.md. | 1 min |
| 3 | Open the assistant's memory settings (Privacy settings pass lists where) and compare the two. | 2 min |
| 4 | Mark each item keep, correct or delete. Delete the wrong, stale and sensitive ones in the settings. | 3 min |
| 5 | Choose a setting for each kind of task (table below), then put the next audit in your calendar, three months out. | 2 min |
List everything you currently remember or infer about me from saved memories
and past chats. Group under: identity and work, preferences, ongoing projects,
people I have mentioned, opinions you think I hold, anything else. For each
item give the memory, roughly when it was learned if known, and whether I
stated it or you inferred it. Output a Markdown list. Do not add anything you
are not using.
The saved list is personal data. Keep it in a private folder, and read What not to paste before you share it with another assistant. If you moved from another assistant, check what came with you. Gemini can import your memories and your whole chat history from ChatGPT or Claude (personal accounts, 18 or over, not in the EEA, Switzerland or the UK). Claude can import memories, not old chats, through a prompt you run in the other assistant and paste back. An import brings stale and sensitive items along, so run steps 1 to 4 on the new assistant after any import.
If a memory is something you want every chat to know, put it in the text you own instead (Tell it who you are, once). When that text and a memory disagree, the text wins.
Which setting for which task
| Situation | Setting we suggest |
|---|---|
| Casual questions and style preferences | Memory on |
| A long project | A project with its own memory and files, and no global memory |
| Health, money, legal, relationships, other people's details | A temporary or incognito chat |
| Debugging an odd answer | A temporary chat, to compare with and without memory |
| Reading untrusted web pages or documents in the chat | Memory off, or review the memories afterwards |
A temporary chat is not the same as nothing stored. Vendors keep temporary chats for a period; Privacy settings pass gives each vendor's stated retention.
How to tell it worked
You have a dated list in a file, every item on the memory page is one you chose to keep, and you know which kind of chat to open for which task.
Signs that memory is hurting again: answers refer to irrelevant past details, the assistant agrees too readily (Getting disagreement instead of agreement), or it "knows" something that is no longer true. When you see one, run the audit early.
For agents (AI tools that take actions for you) that can read your files and email, see Letting an agent near your email, browser and files.
Sources
- PersistBench, arXiv 2602.01146, ICML 2026. Peer-reviewed benchmark, 18 models. Link
- The Hacker News, 2024. Report on a ChatGPT macOS memory flaw, since patched. Security report. Link
- OpenAI. Memory FAQ. Vendor help. Link
- Google. Import memory and chats to Gemini. Vendor help, read 2026-10-18. Link
- Tom's Guide. How to move your ChatGPT memory to Claude, March 2026. News. Link
Changes
- v1.010 Oct 2026first version
- Time to do it
- 10 minutes (our estimate)
- Sources
- 3
- Updated
- 10 Oct 2026
- Last checked
- 18 Oct 2026