Anyone who leans on a chatbot every day should understand what personal data follows them from one session to the next. All three major assistants — ChatGPT, Gemini and Claude — now ship with a memory layer that is on by default or close to it, a dashboard where you can audit it, and a temporary or incognito mode for conversations you would rather not leave behind. The catch lies in the fine print: deleting a chat does not always erase what the assistant learned from it, and a temporary chat is not the same as a chat nobody stores.
The details below come from the official documentation published by OpenAI, Google and Anthropic, reviewed on October 28, 2026. All three warn that features vary by plan, country and app version.
- Memory is a separate layer: It is not real-time model learning but a store of facts injected into context when the assistant deems them relevant.
- Deleting a chat is not enough: In ChatGPT and Claude, removing the conversation does not remove the memory entries derived from it.
- Temporary modes still retain data: ChatGPT keeps temporary chats up to 30 days, Gemini up to 72 hours, and Claude 30 days by default.
What does it actually mean for a chatbot to remember?
It means that when you open a new conversation, the assistant can draw on details you shared earlier without you repeating them. The model is not learning about you in real time; a separate layer stores or locates information and adds it to the context when it seems pertinent.
That layer has several components that are erased independently. OpenAI lists past conversations, saved memories, custom instructions, library files (unavailable in the European Economic Area, Switzerland and the UK) and connected apps such as Gmail. Google refers to previous conversations, connected Google apps and instructions. Anthropic separates search across past chats from memory proper. Custom assistants such as GPTs, Gems and Claude Projects add yet another layer: each Claude project carries its own memory.
ChatGPT: where to look and what does not delete itself
The controls live under Settings, inside Memory or Personalization depending on the app version. Depending on your account you will see toggles such as Reference saved memories and Reference chat history. You can also simply ask the assistant what it remembers about you.
Three documented details deserve attention. First, deleting the chat does not delete the memory: removing the original conversation does not automatically erase the saved memory that came from it, so a full removal requires deleting both the memory and the chat. Second, switching off is not instant: if you disable Reference chat history, what was recalled from past chats is scheduled for deletion from OpenAI’s systems within 30 days. Third, records persist: OpenAI may retain logs of deleted memories for up to 30 days for security and debugging.
The Don’t mention this again option reduces future references to a fact but does not erase its origin. And turning memory off does not disable a limited use of context that OpenAI reserves for high-risk situations.
Gemini: memory rides on saved activity
In Gemini the path is Settings and help > Personal intelligence > Memory. To use it you must be over 18, hold a personal account (not a work or school one) and have Save activity enabled. Google notes that, for now, it is unavailable in Gems and Gemini Live conversations.
To find out whether it has pulled from past conversations, just ask. To delete what it knows about you, remove the conversations containing that detail from Activity in Gemini Apps, though it may take a while to stop using them. If the detail also sits in a connected app, you must disconnect it: doing only one of the two things may leave Gemini still able to find it.
There is small print for European users. Personalization with connected Google apps (Gmail, Photos, YouTube and others) is not available in the European Economic Area, the UK, Switzerland or Nigeria. The Memorized data summary, a single screen showing what Gemini understands about you, is limited to US users on a Google AI plan and only in English. And on asking Gemini to memorize a specific fact, Google’s own help contradicts itself: one page limits it to the US and paid plans, while another explains how to do it from Saved info without country restrictions.
Claude: topic-based memory and a per-conversation switch
Anthropic explains that Claude stores memory as a set of topics while you converse, not as a summary at the end. By default it avoids sensitive topics such as health, ethnicity, religion, politics or gender identity unless you explicitly enable Include sensitive topics in memory.
According to its documentation, memory is on by default on Free, Pro and Max plans; on Team and Enterprise the organization owner controls it. It is managed under Settings > Your Claude: Manage next to Your memory shows the list of topics, which you can read, edit or delete one by one. When disabling it you choose between Pause memory, which keeps what is stored but stops using it, and Reset memory, which wipes everything, including project memories, with no way back.
Search across past chats is a different feature and is limited to paid plans. Claude shares ChatGPT’s caveat: deleting a conversation does not remove the memory entries generated from it; they must be deleted separately.
Temporary or incognito chat: what does it really save you?
It saves you from having that conversation feed memory or appear in your history, but not from the provider storing it for a few days. In ChatGPT’s temporary chat you choose beforehand whether existing memory is used, and no new memory is created; if you later save the chat it becomes a normal one and can then feed memory. Gemini’s temporary conversations create no memory and are kept up to 72 hours. Claude’s incognito chats use no memory and create none, are retained 30 days by default, and cannot be saved afterward. Claude also offers a middle path: turning memory off for a single conversation from the plus menu before the first message, which neither uses nor creates memory but still appears in your history.
The broader lesson is that memory has been marketed as convenience while being engineered as a multi-layered system the user cannot fully see. That two of the three services require deleting the chat and, separately, the memory derived from it is a design that favors data retention. And Gemini’s contradictory help pages on what can be memorized in Europe do little to inspire confidence. A sensible habit is to audit the memory screen periodically, much as you review app permissions, and to use temporary mode for any query you would not want resurfacing months later.