ChatGPT Memory: How It Works, Limits and How to Manage It
Saved memories, chat history, the memory summary, plans, memory full, temporary chats and projects, explained.
ChatGPT memory is the feature that lets ChatGPT carry details about you from one conversation into the next. It works through saved memories and reference chat history, and since June 2026 the newer version keeps itself current and shows you a summary of what it knows.
Below: how each part works, which plans get what, how to view or turn it off, what "memory full" means, and how it compares to Claude's memory.
How does ChatGPT memory work?
ChatGPT memory works by combining facts it stores about you with context it retrieves from your past chats, then using whatever looks relevant to shape each answer. According to the OpenAI Help Center, memory can draw on five sources, depending on plan and region: past chats, saved memories, custom instructions, files in your Library, and connected apps such as Gmail.
The model itself remembers nothing between sessions. Memory is an app layer: ChatGPT keeps notes about you and feeds the relevant ones back into new chats. It doesn't search your history on every message, only "when it is likely to improve a response", so two similar questions can get differently personalized answers.
Most of it happens without you asking. Researchers led by the Max Planck Institute for Software Systems analyzed 2,050 memory entries donated by 80 ChatGPT users and found that 96% had been created by ChatGPT on its own, not because the user said "remember this"; 28% contained GDPR-defined personal data (Dash et al., 2026).
What is the difference between saved memories and chat history?
Saved memories are explicit facts, such as "works in TypeScript" or "vegetarian", that ChatGPT keeps until you delete them. Reference chat history is looser: ChatGPT picks details out of past conversations, and OpenAI warns that what it derives this way "can change as ChatGPT updates what is most useful to remember." Saved memories are stored separately from your chats, and there's no separate storage limit for what ChatGPT can reference through chat history.
The two toggles are linked. Turning off Reference saved memories also turns off Reference chat history, but you can keep saved memories on with chat history off.
What changed in ChatGPT memory in 2026?
In 2026, ChatGPT memory started updating itself and got a readable, editable memory summary. On June 4, 2026, OpenAI began rolling out what it calls a "more capable and scalable system for synthesizing memory" to Plus and Pro users in the US, with Free, Go and more countries following over the next weeks (OpenAI, 2026).
Three things are different in the improved memory:
- Memories update automatically. ChatGPT revises time-sensitive details as they age, so "is going to Singapore in July" becomes a past trip once July is over.
- The memory summary. A page in Settings shows a high-level view of what ChatGPT knows. You can correct it, mark a detail "Don't mention this again", or delete it. Since June 12, 2026, you can delete what appears there and turn memory off from the same page (ChatGPT release notes, 2026).
- More room. OpenAI said it would increase memory capacity for Plus and Pro; 9to5Mac reported that storage doubled.
The summary isn't the full record. OpenAI calls it a high-level view, not a complete list, so ask ChatGPT "What do you remember about me?" to see more. If you prefer the old explicit list, you can switch back to legacy saved memories in Memory settings without losing chat history. Since May 5, 2026, the Sources icon under a response also shows which memories, past chats and custom instructions shaped it.
Which ChatGPT plans have memory?
Every ChatGPT plan has memory, but the features vary by plan, region and workspace. As of October 2026, OpenAI's docs describe it like this:
| Plan | What memory includes |
|---|---|
| Free and Go | Saved memories, past chats and custom instructions as sources; improved memory announced for rollout after Plus and Pro |
| Plus and Pro | Improved memory with the memory summary (US first, from June 2026), more capacity, and Library files and connected apps such as Gmail as sources where your region has them |
| Business, Enterprise, Edu | Admins manage memory settings and permissions; content isn't used for training by default |
| Regulated workspaces | Improved memory off by default in ChatGPT for Healthcare and Enterprise with Regulated Workspace |
Controls such as searching and sorting saved memories, automatic prioritization and version history depend on plan and platform. Region matters as well: when OpenAI shipped chat-history memory in Europe in 2025, it was off by default and had to be switched on in settings (it-daily, 2025).
How do you view, edit or turn off ChatGPT memory?
Open your profile picture > Settings > Personalization > Memory. That one screen holds every memory control: the memory summary and saved memories to read and correct, delete options, and the two toggles. Turning off Reference chat history schedules what ChatGPT learned from past chats for deletion within 30 days, while the chats themselves stay in your sidebar.
Deleting a saved memory doesn't delete the chat it came from, and deleting a chat doesn't necessarily delete the memory, so a real cleanup touches both. The steps are in how to delete ChatGPT memory, and the setup side, from saving a fact to sealing off a project, is in how to save context in ChatGPT.
Why does ChatGPT say memory is full?
ChatGPT says memory is full when the saved-memory list reaches its storage cap, a size OpenAI has never published. Users report the notice as "Memory is full. New memories won't be created" (OpenAI Community), and from then on new facts stop sticking, even when the chat says "Memory updated".
It mostly affects the legacy saved-memory list. Depending on plan and platform, saved-memory controls include automatic prioritization, which helps avoid the full state, and the improved memory keeps track of "the details it determines are most important" instead of a fixed list of items (ChatGPT release notes, 2026).
If you hit it, in this order:
- Delete what's no longer true: old jobs, finished projects, one-off trip details.
- Merge near-duplicates. Delete five memories about your writing style and ask ChatGPT to save one that covers all of them.
- Move standing rules ("short answers, no preamble") into custom instructions. They're a separate setting and don't take up saved-memory space.
- Keep Reference chat history on. It has no separate storage limit, though it's less precise.
What does temporary chat do to memory?
A temporary chat never creates or updates memories, and it doesn't appear in your history. Since August 27, 2026, you choose before your first message: a personalized temporary chat can use your existing memories, custom instructions and plugins, and a non-personalized one uses none of them (ChatGPT release notes, 2026). You can't change the choice once the chat starts, and OpenAI may keep a copy for up to 30 days for safety (OpenAI Help Center).
How does memory work in ChatGPT projects?
Each project uses either default memory or project-only memory. With default memory, project chats can use your saved memories. With project-only memory, saved memories stay out, and chats can reference other conversations in the same project but nothing outside it (OpenAI Help Center).
Since August 14, 2026, eligible unshared projects can switch between the two under Project settings; shared projects are always project-only (ChatGPT release notes, 2026). Project-only memory is the right setting for client work you don't want leaking into everything else.
What are the limits of ChatGPT memory?
ChatGPT memory has four limits worth planning around:
- It's inferred. Most memories are ChatGPT's own interpretation of what you said, so they can be wrong, stale or oddly specific. Check the summary now and then.
- It's not guaranteed. On November 6, 2025, OpenAI opened an incident for missing memories and could restore them only partially for some users (OpenAI Status, 2025).
- It only knows what you typed. The client thread in Gmail (unless connected), the plan in Notion, the decision from last week's call: none of it is in memory unless you said it in ChatGPT.
- It stays in ChatGPT. Claude, Cursor and Codex can't read it, so every other tool starts from zero.
ChatGPT memory vs Claude memory
ChatGPT memory is more automatic; Claude memory is more transparent. Neither one shares anything with the other.
| ChatGPT | Claude | |
|---|---|---|
| How it's stored | Saved memories plus a memory summary that updates itself | Individual topics saved as you chat |
| Past chats | Reference chat history (no separate storage limit) | Chat search on paid plans |
| Where to manage it | Settings > Personalization > Memory | Settings > Your Claude (Settings > Memory in the older layout) |
| One chat without memory | Temporary chat, personalized or not | Incognito chat, or memory off in the "+" menu |
| Per-project memory | Default or project-only | Each project has its own; "Use account memory" toggle |
| Business plans | Admins manage settings | Owners enable it; off per member until they turn it on |
Sources: OpenAI Help Center, Anthropic Help Center, checked October 2026.
Claude lets you open every topic and edit it, while ChatGPT gives you a summary and a list. Claude can also import memory from ChatGPT with a copy-paste prompt; the details are in the Claude memory guide, and the full switch is in how to move context from ChatGPT to Claude. For Gemini in the mix, see which AI has the best memory.
When is ChatGPT memory not enough?
ChatGPT memory isn't enough once you work in more than one AI tool, or once the context you need lives in your apps rather than your chats. Memory remembers what you discussed. It doesn't know who you are across Gmail, Calendar, Notion and Slack, and it can't follow you into Claude or Cursor. That gap is personal context, and it's a different layer from memory.
Unabyss builds that context from the apps where it already lives, keeps it current, and serves it to ChatGPT through an official plugin and to Claude and Cursor over MCP. In an August 2026 Unabyss benchmark, adding it to an assistant that already had connected tools cut cost by 36% and time by 38% (What Structured AI Context Is Worth). Keep ChatGPT memory on for "what we talked about", and compare the cross-tool options in Best AI Memory Tools Across ChatGPT, Claude & Cursor.
