949 words | Esment
Making ChatGPT remember.
What ChatGPT’s memory really is: a summary, saved memories and a look back at old chats. How to turn each on, what it keeps, and when it falls short.
╌╌╌╌
ChatGPT can already use what you said in earlier conversations, as long as Memory is on. Open Settings → Personalization → Memory and make sure it is. If ChatGPT still behaves as if every chat were new, the reason is usually not a missing switch but how its memory is built. It is a set of personalization features that decide for themselves what to bring back, not a store you can query.
What ChatGPT’s memory is made of
OpenAI’s help page lists the sources memory can draw on: past chats, saved memories, custom instructions, files in your Library and, on some plans, connected apps such as Gmail (OpenAI help). Three of them do the remembering:
| Memory summary | Saved memories | Reference chat history | |
|---|---|---|---|
| What it is | A high-level summary ChatGPT keeps updating | An explicit list of items | Context pulled from past chats |
| How it gets there | Written by ChatGPT from your chats | You ask, or ChatGPT saves one | Chosen by ChatGPT when it seems relevant |
| You can read it | Yes, and correct it | Yes, item by item | Only as Sources under an answer |
| Size limit | Not stated | Yes, it can fill up | None, per OpenAI |
The newer experience, which OpenAI calls improved memory, centres on the summary: it “continually updates a broader summary of relevant context”. The older one, now called legacy saved memories, is the explicit list. Where your account allows it, you can switch between them in the same settings screen.
Why it is not a memory system you control
OpenAI is candid about the trade-off. Memory “does not retain every detail from every conversation”, ChatGPT “decides which available information is relevant to a response”, and it does not search your history for every reply, only “when it is likely to improve a response”. What it derives from past chats “can change as ChatGPT updates what is most useful to remember”.
Put together, that is personalization: context assembled for you, chosen by the model, summarised in its words. It is very good at making replies feel familiar. It is not designed to be exact. You cannot ask it for every decision you made about a project, see when a fact was learned, or rely on a specific detail being brought back the next time it matters.
How to make ChatGPT remember more
- Turn on Memory, and Reference chat history if your settings show it.
- State the fact plainly and ask: “Remember that I work in Central European Time.” A saved memory is the most direct route.
- Put the few things that must apply to every reply in custom instructions.
- For a body of work, use a project, so related chats and files stay together.
- Check the memory summary now and then, and correct what is wrong. “Don’t mention this again” only hides a detail; it does not delete it.
- When an answer misses something, ask directly: “What do you remember about my move to Lisbon?”
Where it falls short
- The saved-memories list has a cap, and when it is full nothing new is saved. ChatGPT memory full explains what to do.
- Recall is chosen by the model, so the same question can bring back different context on different days.
- There is no history of a fact. When something changes, the old version is not kept alongside the new one for you to check.
- It stays inside ChatGPT. Claude, Cursor and every other assistant start from nothing.
When to add a separate memory
If ChatGPT’s own memory covers you, keep using it. Add a separate one when you need exact recall, a history you can inspect, or the same memory in more than one assistant.
Esment is our product, and this is the job it does. ChatGPT connects to it as a custom MCP connector. Every memory is stored as its own entry with its type, its tags, when it was saved and which assistant saved it, and search works by meaning, so you can ask for “everything we decided about pricing” and get the entries, not a paraphrase. Facts that change keep their earlier versions. ChatGPT does not read Esment’s instructions by itself, so you teach it once to check memory before answering. Setup is on the ChatGPT page, and the Mac app can import your ChatGPT history to start from what you already said.
Common questions
Does ChatGPT remember across chats?
Out:Yes, when Memory is on. It can use its memory summary, your saved memories and, if Reference chat history is enabled, relevant parts of past conversations. It chooses what to bring back, so it will not recall every detail.
Why did ChatGPT forget something I told it?
Out:Usually because it was never saved as a memory, the chat was temporary, ChatGPT did not judge it relevant to the new question, or the saved-memories list is full. Ask it directly what it remembers, and save important facts explicitly.
Does deleting a chat delete the memory?
Out:No. Saved memories are stored separately from chats. Delete the memory in settings and the chat where you shared it.
Can Claude or Cursor use my ChatGPT memory?
Out:No. ChatGPT’s memory belongs to your OpenAI account. To share memory between assistants, connect them to the same external memory, as our guide to sharing memory between ChatGPT and Claude explains.
Written by the team that builds Esment, so we are not a neutral party. Tools change quickly: check each one before you rely on it.
Get the next chapter by email.
One email per new chapter. Nothing else.