Snapshot caveat: Memory features change fast. Rollouts, defaults, plan availability, and setting labels move constantly, so check each app's own settings before trusting a path here. Three are mid-flight, ChatGPT's rebuilt auto-updating memory, Claude's free-tier memory since March 2026, and Gemini's Personal Intelligence beta. Reflects June 2026.
AI Memory and Personalization
Using AI
9 min read
In Short
Memory is your chat app keeping a profile of you and drawing on past conversations so you stop repeating yourself. ChatGPT, Claude, and Gemini all do this now, each in a different place and with a different default, and Gemini's is on by default. It is useful and it is a risk at the same time, because a wrong or stale fact, or in documented attacks a planted one, can quietly shape later answers. You can view, edit, delete, and switch it off in every app, and each has a no-memory mode, so the practical thing is to know where your switch is.
01. What It Is
Memory bundles two separate things under one word. One is a stored profile of facts and preferences about you. The other is the app drawing on your past conversations. Without it a chatbot is "stateless," starting each new chat from zero with no idea who you are or what you said yesterday. Memory makes it "stateful." OpenAI labels its two controls "saved memories," what you asked it to keep, and "reference chat history," insights it gathers on its own.
Memory is not training and not custom instructions.
Training improves the model for everyone and has its own opt-out, covered in ai-privacy-and-your-data. Memory personalizes only your answers. Custom instructions are a note you write by hand, such as "answer briefly, I am a teacher," while memory is what the app stores for itself.
See how-to-use-an-llm.
02. Why It Matters
The benefit is continuity. Tell the app once that you are vegetarian, write in British English, or are planning a trip to Tokyo, and it carries that forward instead of asking again. OpenAI's own examples are remembering your recipe preferences or how you like meeting notes formatted.
The same continuity is the risk surface. A fact stored wrong, or one no longer true, keeps shaping answers until you find and remove it, and personal details can surface at an awkward moment. Because each app puts the controls in a different place with a different default, what matters is knowing where the switch is and how to see what the app thinks it knows.
03. How It Works
Which plan unlocks memory differs by app, see chat-apps-and-subscriptions.
What gets remembered
The stored profile is a running set of facts and preferences. Claude keeps work-relevant context like your role, projects, and coding preferences. ChatGPT mixes facts you saved with insights from your chat history. Gemini goes furthest, also drawing on connected Google apps such as Gmail and Photos.
ChatGPT
ChatGPT memory works two ways. "Saved memories" are things you explicitly asked it to keep. "Reference chat history" lets it draw on insights from all past conversations on its own. It launched as a test on 13 February 2024, became general in September 2024, gained cross-chat referencing on 10 April 2025, and reached free users on 3 June 2025.
OpenAI is rolling out a rebuilt system with an editable "memory summary" that auto-updates, starting with Plus and Pro in the US. You read it in Settings under Memory and edit it by typing, or highlight a line and choose "Don't mention this again," which suppresses that detail without deleting it.
Saved memories sit in a separate "notepad," apart from your chats. Deleting a conversation does not remove a memory taken from it, and deleting a memory does not remove the chat where you said it.
Claude
Claude splits the feature in two. "Search and reference chats" is retrieval you trigger, where you ask it to look something up and it runs a search shown as a tool call (see rag). "Memory" works on its own, a running summary of you that refreshes every 24 hours and feeds each new conversation. Each Claude Project keeps its own separate memory space, walled off from other projects, which is Claude's built-in answer to context leakage.
You control it in Settings under Capabilities. "View and edit memory" shows everything it remembers. "Pause memory" keeps it but stops using and adding to it, and "Reset memory" permanently deletes all of it, including project memories, with no undo. Memory reached Pro and Max on 23 October 2025 and free accounts in March 2026, alongside an experimental import and export between Claude and other AI tools.
Gemini
Google calls its personalization "Personal Intelligence," drawing on three sources, your past chats, content from connected Google apps, and any instructions you saved. For each prompt it decides whether these would help.
Gemini's memory of past chats is on by default. Announced on 13 August 2025 alongside Temporary Chat, it requires you to be 18 or older, signed in with a personal Google Account rather than a work, school, or supervised one, with "Keep Activity" on. You turn it off in Settings under Personal context at "Your past chats with Gemini."
Gemini has no separate memory list. To remove what it knows, you delete the past chats themselves in Gemini Apps Activity, and anything pulled from a connected app needs both the chats deleted and the app disconnected. You can ask "Did you use any info from past chats?" to check.
Other assistants such as Copilot, Grok, and Meta AI are converging on the same pattern, an easily enabled memory with a settings switch, not verified per app here.
How to view, edit, delete, and turn it off
This is the section to bookmark. The controls sit in a different place in each app.
| App | Where memory lives | See it | Delete one item | Turn it all off | No-memory mode |
|---|---|---|---|---|---|
| ChatGPT | Settings, Personalization (Memory) | Settings, Memory, or Manage memories | Remove the saved memory in Manage memories, separate from the chat | Toggle off saved memories, which also turns off reference chat history | Temporary Chat |
| Claude | Settings, Capabilities | View and edit memory | Edit the summary, or delete the cited past chat | Pause (keep but stop using) or Reset (delete all) | Incognito chat |
| Gemini | Personal context, or Personal Intelligence | Ask "did you use info from past chats?", or open Gemini Apps Activity | Delete that chat in Gemini Apps Activity | Switch off "Your past chats with Gemini" | Temporary Chat |
Temporary and incognito chats
Every app has a mode that skips memory. ChatGPT's Temporary Chat, the "Temporary" button at the top right, uses no existing memories, creates none, stays out of history, and is not used for training, though OpenAI may keep a copy up to 30 days. Claude's Incognito chat, the ghost icon on all plans including free, is not saved to history or memory, not used for training, cannot be reopened, and is kept 30 days by default. Gemini's Temporary Chat is kept up to 72 hours and not used to personalize or train.
These modes leave your memory untouched and keep the chat out of history and training. That part is real, but they are not anonymous. The conversation still reaches the company's servers and is kept for a short safety window, so they hide the chat from your assistant, not from the provider.
See ai-privacy-and-your-data.
04. Key Terms
| Term | Plain meaning |
|---|---|
| Memory (stateful vs stateless) | Whether the app carries what it knows about you between conversations. Stateless forgets you every new chat, memory makes it remember. |
| Saved memory / memory summary | The stored profile of facts and preferences about you, your "notepad" in ChatGPT, "memory" in Claude. You can usually view and edit it directly. |
| Reference / search past chats | The app pulling from your earlier conversations. ChatGPT and Gemini do this automatically, Claude does it automatically for memory but only on request for search. |
| Personalization / personal context | The umbrella term, Gemini's "Personal Intelligence," for tailoring answers to you, drawing on past chats, saved facts, and sometimes connected apps like Gmail. |
| Custom instructions | A standing note you write by hand about how you want the AI to respond. Different from memory, which the app builds on its own. |
| Temporary / Incognito chat | A no-memory mode. The conversation is not saved to history, does not use or update memory, and is not used for training. It still reaches the company and is kept briefly for safety. |
| Memory poisoning | When hidden instructions, for example on a web page you ask the AI to summarize, get written into its memory and quietly affect later answers. A documented prompt-injection attack. |
05. Examples
- You told ChatGPT once that you are vegetarian, and weeks later it still steers recipes away from meat without being reminded. That is memory doing its job.
- An answer comes back oddly personalized or subtly wrong. Before blaming the model, check the memory for a stale fact it is still applying.
- Gemini refers to something you never told it directly. That is on-by-default memory plus connected Google apps. Check Personal context and Gemini Apps Activity to clear it, or use Temporary Chat for a one-off you do not want stored.
06. Common Misconceptions
"Deleting the chat deletes what it learned."
Not in ChatGPT, where saved memories sit in a separate notepad, so you delete the memory item as well as the chat. In Gemini it is the reverse, deleting the chat is how you remove the memory, after a short delay. The apps differ in practice.
"Turning memory off erases what it already knows."
Off is not delete. ChatGPT's saved memories stay stored when switched off, and you delete them separately. The exception is "reference chat history," which when off does delete what it derived from past chats within 30 days. Claude splits this into "Pause" versus "Reset."
"A saved memory is always correct."
A memory can be stale, a one-off now treated as a standing preference, or self-contradictory. OpenAI rebuilt its system because old memories went stale and clashed, its example being "I am training for a marathon" plus "I sprained my ankle." Applied silently, a wrong memory can skew answers for a long time, so check it before blaming the model.
"Temporary or Incognito chat is fully private and anonymous."
It skips memory, history, and training, which is real and useful, but it is not anonymous. The conversation still goes to the company's servers and is kept for a short safety window, about 30 days for ChatGPT and Claude and 72 hours for Gemini, and may be reviewed for abuse.
"A memory only ever holds things I told it."
Not guaranteed. In November 2025, Tenable's "HackedGPT" research showed that asking ChatGPT to summarize a booby-trapped web page could plant hidden instructions into its memory, which then persisted and re-executed in later conversations. This is prompt injection writing a false memory into your account, the clearest reason to be able to see and clear what your assistant remembers.