As of August 2026, ChatGPT, Claude, and Gemini each retain context across sessions by default for most paid and free users, but the three systems differ in what they store and how easy they are to switch off. Per OpenAI's Memory FAQ, ChatGPT keeps a synthesized summary of past chats; per Anthropic's help center, Claude's new memory experience focuses on "role, projects, and professional context"; per Google's Gemini Apps Help, Gemini's memory requires an 18+ personal account with Activity history turned on.
The three vendor help pages are the only sources for the claims below, and each is dated or marked as recently updated at the time of writing. Nothing here is a benchmark or a lab test — it's a reading of what each company tells its own users, translated into what it means for someone using these tools at a desk.
What does ChatGPT's memory actually store?
Per OpenAI's Memory FAQ, ChatGPT's current memory system automatically builds "a high-level view" of useful details from conversations, files, and connected apps, rather than logging every line you type. OpenAI notes this replaced an older "saved memories" system, which required you to explicitly tell the model to remember something and which the company says "often became stale." Both systems still exist: you can switch back to the legacy saved-memories view under Settings, and OpenAI is explicit that deleting memories does not delete your underlying chat history.
A separate question — whether your chats train the underlying model — is handled by a different toggle. Per OpenAI's Data Controls FAQ, ChatGPT uses conversations to improve its models by default, and turning off "Improve the model for everyone" stops that without touching your chat history or memory. Temporary chats are excluded from training automatically and are deleted after 30 days, per the same page.
What does Claude's memory cover, and who gets it?
Anthropic splits this into two related features. Chat search, which lets Claude look back through your conversation history, is available on Pro, Max, Team, and Enterprise plans across web, desktop, and mobile, according to Claude's help center. Memory — the feature that carries context forward into new chats without you re-explaining it — is rolling out on free, Pro, and Max plans in a new form, while Team and Enterprise accounts remain on an older, separate version with its own timeline, per the same article.
What Claude's memory is scoped to track, per Anthropic's documentation, is deliberately narrow: role and project context, communication preferences and working style, and technical or coding preferences. You can pause memory, which stops new entries while keeping what's already stored, or reset it entirely, which Anthropic describes as an irreversible deletion. Incognito chats sit outside memory altogether, on any plan.
How is Gemini's memory different?
Google's version, per its Gemini Apps Help page, is gated more tightly at the account level than the other two: it requires a personal Google Account, not a work or school one, a minimum age of 18, and the account-wide "Keep Activity" setting switched on. Where ChatGPT and Claude build a running internal summary, Google frames Gemini's memory as reading directly from your Gemini Apps Activity log — which is also how you audit or delete it.
Google's help page includes a detail the other two don't: you can ask Gemini directly whether a given answer drew on a past chat, and it will tell you. Turning the feature off happens through Settings & Help → Personal Intelligence → Memory. Deleting the underlying record, though, is a two-part job — Google's documentation notes that disconnecting a linked app isn't enough on its own if the relevant chats are still sitting in your Activity log; you need to remove both.
How do the three compare, side by side?
| Feature | ChatGPT | Claude | Gemini |
|---|---|---|---|
| What's stored | Synthesized summary of chats, files, connected apps | Role/project context, working style, technical preferences | Details drawn from Gemini Apps Activity log |
| Where it lives | Settings > Personalization > Memory | Settings > Memory (new) or Capabilities (legacy) | Settings & Help > Personal Intelligence > Memory |
| Off switch | Toggle off; delete individual memories via three-dot menu | Pause (stops new entries) or reset (deletes all, irreversible) | Toggle off, plus manual cleanup of Activity log |
| Separate from chat history? | Yes — deleting memory does not delete chats | Yes — memory and chat search are distinct settings | No — memory reads from the Activity log directly |
| Excluded mode | Temporary chats (not saved, not used for training) | Incognito chats (no memory access) | Not documented on the cited page |
Does turning off memory also stop model training?
Not automatically, and this is the distinction most people miss. For ChatGPT, memory and training are two separate settings, per OpenAI's Data Controls FAQ — switching off "Improve the model for everyone" does not delete memory or chat history, and switching off memory does not opt you out of training. Claude's and Gemini's cited help pages address memory and chat retention but don't spell out an equivalent training toggle in the pages consulted for this piece, which is itself worth noting rather than assuming away.
What none of the three vendor pages establish is how long a "paused" or "disconnected" memory persists on the backend before it's actually purged, or whether enterprise admin controls override an individual's personal toggle. Anyone handling client or employer information through these tools should treat the account-level settings described here as the documented floor, not a guarantee, and check their organization's own data policy before assuming a personal toggle is the last word.
For a related business news perspective, read Is AI actually taking entry-level jobs? Here's what the data says.
For more context, read How to set up a passkey and stop relying on passwords.