What ChatGPT's memory is made of
According to OpenAI's help page, ChatGPT draws on several sources: past chats, saved memories, your custom instructions, and depending on plan and region, files in your Library and connected apps such as Gmail. There are two versions. The newer memory keeps a continuously updated summary that you can read and correct. The older one keeps a list of individual saved memories. What you get depends, in OpenAI's words, on plan, region, platform and workspace. You manage all of it under Settings → Personalization → Memory.
“Memory full”: two messages, two causes
- Saved memories full. The list of saved memories used to have a hard cap, and at some point nothing more went in. Since October 2025, ChatGPT sorts for itself: memories you haven't needed in a while move to the background, so the account doesn't hit a “memory full” state. That came first to Plus and Pro on the web. In June 2026 OpenAI also doubled the capacity for Plus and Pro. OpenAI gives no fixed number. You can still sort memories yourself, prioritise or deprioritise them, and restore earlier versions.
- “Chat memory full” in a single chat. This message means something else: according to OpenAI's free-plan FAQ, the context window of the current conversation is full, meaning how much ChatGPT can take into account in that one chat. The fix is a new chat. You don't need to delete any memories for it.
Cleaning up and deleting
- Individual memories: the memory settings show what's stored, and you can delete single entries. Below an answer, ChatGPT also shows which sources it used, and you can correct or delete a memory right from there.
- A chat isn't the memory: deleting a chat doesn't necessarily delete the memory created from it. To get rid of something completely, OpenAI says to delete both the saved memory and the original chat, and to check the summary, archived chats and Library files.
- Retention: OpenAI may keep logs of deleted saved memories for up to 30 days.
- No trace: according to OpenAI, a Temporary Chat neither creates nor updates memories while it stays temporary.
Projects with their own memory
You can set a ChatGPT project to “project-only memory” (Project settings → Memory). Chats in that project then only draw on other chats in the same project, not on the rest of your memory, and nothing flows out. Shared projects always work this way. According to OpenAI's projects help, there's no list of what the project has remembered. Files per project: 5 on Free, 25 on Go and Plus, 40 on Pro, Business, Enterprise and Edu.
Where it stops: three limits
- It stays with OpenAI. ChatGPT's memory lives in your OpenAI account. Claude, Cursor and Mistral Vibe see none of it. OpenAI doesn't say whether saved memories are part of the data export, and when you move exported chats to another account, memories don't come along, per OpenAI.
- Documents only at the edge. Library files can serve as a source, but files and Gmail only on Plus and Pro. The sources below an answer name a chat, a memory or a file. OpenAI doesn't describe citations down to the page, and says itself that sources may not show every factor behind an answer.
- ChatGPT decides what stays. The summary updates itself, and less-used memories move to the background. For preferences that's convenient. For a decision you need to find again in six months, it's an uncertain place.
How to make your knowledge last
What should last is best kept outside ChatGPT, in a memory you control and that every AI app can read. ChatGPT connects to one as a custom MCP server. MCP is the open standard for this, and Claude, Mistral Vibe and Cursor speak it too. Memexa is such a memory: notes, PDFs and scans become a linked knowledge graph, scans run through OCR, and answers cite the source down to the page. Connecting takes one URL:
ChatGPT (chatgpt.com)
chatgpt.com/plugins → Add → Add custom MCP server → name, https://brain.memexa.ai/mcp, OAuth → Create as plugin
Claude (claude.ai)
Settings → Connectors → Add custom connector → https://brain.memexa.ai/mcp
We've tested this path ourselves on the Plus plan. OpenAI's help page so far lists custom MCP servers only for Business, Enterprise and Edu, plus Pro with read permissions, and setup works in the browser only. We haven't tested the free ChatGPT plan. Claude allows one custom connector on its free plan too. Sign in once in the browser, no tokens to copy. The full walkthrough for both apps is in Give ChatGPT and Claude the same memory.
Important: saving doesn't happen by itself
An MCP memory doesn't record your chats. ChatGPT gets tools for searching, reading and saving, and decides during the conversation when to use them. Reading usually works unprompted once a question sounds like it's about your knowledge. Saving happens when you ask: “Remember in Memexa …”. Before changes, ChatGPT asks for confirmation depending on your settings, and under Settings → Plugins → Permissions you decide how often. Documents don't come in through the chat anyway: you upload them to Memexa.
| Where it lives | What it remembers | Citation down to the page | |
|---|---|---|---|
| ChatGPT memory | Your OpenAI account | A summary or list drawn from your chats, files depending on plan | No |
| Project memory in ChatGPT | A single project | Chats and files in that project | No |
| Memexa | Hosted in the EU, in any AI app that speaks MCP | Notes, documents and scans as a linked knowledge graph | Yes |
Where your data lives
Memexa hosts in the EU under the GDPR, your content is never used for training, and everything exports with one click, notes as Markdown. When ChatGPT calls a tool, it gets only the results for that request, not your whole memory. Those results then become part of the conversation at OpenAI. Every provider, with seat and purpose, is listed on the Trust page.