ChatGPT started remembering people across separate chats

Start here

A person mentions once, in passing, that they are vegetarian, and weeks later a recipe suggestion in an unrelated chat quietly avoids meat. That continuity is memory: a feature that lets an assistant carry details from one conversation into another, rather than starting from nothing each time. OpenAI introduced this for ChatGPT on 13 February 2024, in a post stating that remembering things a person discusses across chats saves them from repeating information, and stressing from the start that a person is in control of that memory, with settings to view, correct or turn it off.

What the documents say

The feature has expanded since that first post. OpenAI's original announcement records its own later updates inline: by 10 April 2025, it states, memory works two ways, saved memories a person has asked it to remember, and chat history, insights gathered from past chats to improve future ones, a broader scope than the initial version. The current Memory FAQ describes a memory summary a person can read, edit, or delete outright, noting it will not include everything ChatGPT remembers, so it is a partial window onto a broader synthesis. Google documents a comparable feature for Gemini: its personalization help page states a person can ask directly whether the assistant used info from past chats, and that turning memory off uses a settings toggle, with full removal requiring deletion of the underlying chats from Gemini Apps Activity, not just the memory feature.

Check this

Both vendors document a direct way to audit memory rather than guess at it: ask the assistant what it remembers or whether it used saved information to answer a specific question. A reader concerned about a fact lingering can also check the more demanding requirement each vendor states for erasing it: OpenAI's FAQ says a memory can outlive the chat that created it, so deleting a conversation does not remove it, and Google's page says removing source chats from Activity may still leave a short delay before the information stops shaping responses.

What holds and what fails

The core claim, that these systems now carry details across sessions rather than forgetting at the end of every chat, holds as a documented product change, dated from a specific 2024 launch and expanded since. What the documentation does not offer is a promise that turning memory off retroactively erases everything already said: past chats containing a now-deleted memory remain in history unless deleted separately, on both platforms. Treating memory as off with the assistant having forgotten everything it was told is the gap between the setting and the archive underneath it.

• Ask the assistant directly what it remembers, rather than assuming a setting shows the full picture.
• Check whether turning memory off actually deletes existing memories, or only stops creating new ones.
• Delete the underlying chat history, not just the memory feature, to remove a specific detail for good.

Memory turned a stateless tool into something closer to an ongoing relationship, and the vendors' own pages are explicit that the controls are partial: visible, editable, but not a single switch that erases the past.