What AI assistants remember about the people you mention

Anthropic's help page for Claude lists what its memory is meant to keep. Your role and projects are on the list, and so are "the people and places in your work and life" [1]. That line describes how people actually use assistants: you ask for help with a hard email to a manager, a birthday plan for a sister or a worry about a friend, and the assistant may keep a note about that person as well as about you.

Where that note ends up, and what it takes to remove it, differs between ChatGPT, Claude and Gemini. In ChatGPT and Claude, deleting the conversation doesn't necessarily remove a memory built from it. When you switch assistants, the import prompts the vendors provide ask for family details and relationships, so notes about other people travel unless you edit them out first. This article covers details about other people. For correcting wrong facts about yourself in ChatGPT, see our earlier guide to correcting AI memory.

How often other people turn up

Two analyses of real users' ChatGPT memories give a sense of scale. Both cover ChatGPT only.

The larger one is Kiran Garimella's independent analysis, published on Substack in June 2026. It covers 12,112 memories from 766 users, taken from about 1,200 donated exports. Of those memories, 703 (5.8%) contained identifying information about real people other than the user. Examples include children's names and ages and spouses' employers. Those memories came from 205 users, just over a quarter of the people with any memory stored [2].

A second figure from the same post gets quoted more often: names appear in 1,770 memories (14.6%). That count includes any name, and the analysis does not separate users' own names from other people's. Only the 5.8% figure is about third parties. The method also needs a caveat. GPT-5 applied the labels across nine categories, the author says the results "may not be fully correct", and the post was not peer reviewed [2].

The peer-reviewed study is smaller: Dash and colleagues (The ACM Web Conference 2026) examined 2,050 memory entries from 80 users. They found 96% had been created by the system on its own rather than on request, and 28% contained personal data as defined by the GDPR [3]. The paper doesn't count other people's details separately, so it can't answer the third-party question on its own. It does support the practical point: most memories are written without you asking.

Where the detail sits in each assistant

Every cell below comes from the vendor's help page. "Not documented" means the page doesn't say. Controls vary by plan, region and app version.

Sources: OpenAI [4], Anthropic [1], Google [5].

What deleting actually removes

In ChatGPT, a saved memory is stored separately from the conversation that produced it, which is why OpenAI lists several places to clear, as in the table above. "Don't mention this again" only reduces future references; it deletes nothing. OpenAI may keep logs of deleted saved memories for up to 30 days. If you turn off Reference chat history, details remembered from past chats are scheduled for deletion within 30 days [4].

In Claude, the limit is stated more firmly. Anthropic says that when a conversation expires or is deleted, the memory entries generated from it stay until you delete them yourself. Deleting an old chat about a friend therefore leaves the topic about that friend in place. Open Settings > Memory, find the topic and delete or edit it. "Reset memory" permanently deletes all memory, including project memory. "Pause memory" keeps everything and stops Claude from using or adding to it [1].

In Gemini, the chats are the memory's source, so you remove a detail by deleting every chat that contains it from Gemini Apps Activity. Google notes there may be a short delay before Gemini stops using a deleted chat. If the detail came through a connected app, you also have to disconnect that app; Google says both steps are required [5]. Instructions you typed for Gemini yourself are separate, and you can edit or delete them one at a time [6]. Gemini Apps activity older than 18 months is deleted automatically by default. You can change that to 3 or 36 months or turn it off [7].

What moves when you switch assistants

Memory import works by asking your old assistant to write out everything it knows, then pasting that text into the new one. Anthropic's recommended prompt for this asks the old assistant for "Personal details: name, location, job, family, interests". It also tells it: "Do not summarize, group, or omit any entries." Anthropic says you can edit the prompt to leave out sensitive information. It adds that Claude focuses on work-related topics and "may not retain imported personal details unrelated to work", and calls imports experimental [8].

Google's announcement says Gemini's import brings over "key preferences, relationships, and personal context", and gives "your sibling's name" as an example [9]. Google's privacy hub says imported memory and chats are saved in your Activity and used to improve its services, including training generative AI models [10]. The import isn't available in the EEA, Switzerland or the UK, or for work, school or supervised accounts [11].

OpenAI's export page says the ZIP "includes your chat history and other relevant account data". It doesn't say whether saved memories or the memory summary are part of that [12].

For a memory import, the text you paste is the text that gets imported. Read the block before you paste it and delete the lines about other people you don't want carried over. Gemini can also import your full ChatGPT or Claude chat history from an export ZIP; that brings every conversation as you wrote it, including what you said about other people, so delete those chats before exporting or skip the history import. [11] If Claude is the assistant you're leaving, Anthropic suggests asking it to "Write out your memories of me verbatim" to see what it holds [8].

Before you name someone

This is an editorial rule, based on the controls above.

1. Decide whether the name is needed. "A colleague who prefers morning meetings" usually gets you the same help as a full name.
2. Keep other people's sensitive matters out of a memory-on chat. That covers health, money, relationship trouble and legal problems. Claude never saves a few categories, such as government ID numbers and criminal history, and by default doesn't store topics like health or religion [1]. Those rules apply to kinds of information, whoever it belongs to. A friend's name, job or relationship to you falls outside them.
3. Choose a no-memory chat before you start. A ChatGPT temporary chat doesn't create or update memories, and choosing Unpersonalized also keeps it from using existing ones; that choice can't be changed once the chat starts [4]. Claude's incognito chats aren't saved to memory or history and are kept for 30 days by default [13]. Gemini's temporary chats can't save information for personalization and are kept with your account for 72 hours [14] [10]. None of these means nothing is kept.
4. Audit by searching. Look through your memory list for relationship words such as sister, son, boss and ex, as well as for names.
5. Delete in both places. Remove the memory entry, then the chat or chats it came from.

A worked audit (fictional)

Noor, an invented person, is moving from ChatGPT to Claude. In her saved memories she finds six entries that mention someone else. All six are invented for this example.

One entry kept, two rewritten, three deleted. Noor then runs the export prompt. Before pasting the output into Claude, she deletes any lines that repeat the three deleted entries, since chat history can bring them back. Afterwards she opens Settings > Memory in Claude to check what was actually saved.

A note on Kluro

Kluro is a Mac app built around relationship memory, so it keeps details about other people by design. It builds them from the conversations you connect. Its saved memory stays encrypted on your Mac, and enabled hosted AI processes relevant content for answers and background understanding. You can disconnect a source to stop syncing, and removing that source's imported data also deletes its saved memory. The privacy page sets out those controls, and How Kluro uses AI explains how corrections work.

How this was checked

We read every help page cited here in full on 30 September 2026. OpenAI's help pages refused automated requests that day. We therefore read the memory page from an Internet Archive copy saved on 24 September 2026, five days after the page's last listed update. The export page came from a copy saved on 31 August 2026. We did not test any assistant with a real account.