Why does ChatGPT remember some things but not others?

The model decides what is worth writing to memory, and it is conservative. Stable facts about who you are tend to stick. Details specific to one task usually do not. You can force something in by stating it plainly, for example "Remember that I…", but you are competing for a small number of slots.

Can ChatGPT read my previous conversations?

It can reference recent chat history in some configurations, but that is not the same as remembering. Referencing is a lookup at the provider's discretion; memory is a durable stored fact. Neither gives you a model that has genuinely accumulated two years of knowing you.

Does starting a new chat erase what ChatGPT knows?

It erases the working context — everything the model could see inside that thread. Your saved memories survive. That is why a new chat can feel like meeting a colleague who read your one-paragraph bio but was not in any of the meetings.

How do I make ChatGPT remember more across conversations?

You cannot enlarge the memory bucket. The three things that work are: curate the memory panel so the slots hold what matters, paste a compact personal summary at the top of important threads, or keep the full version in an external vault and connect it to ChatGPT so the detail is available on demand rather than crammed into memory.

A field note · ~5 min read

Does ChatGPT remember across chats?

Q: Does ChatGPT remember across chats?

Partly. ChatGPT carries a small set of saved memories — typically 8 to 12 short facts — into every new conversation. Everything else you said lives only inside the thread where you said it and is gone once that thread ends. So it remembers a sketch of you, not the conversations themselves.

Partly — and the part it keeps is smaller than almost anyone expects. A short list of saved facts follows you into every new conversation. The conversations themselves do not. Knowing exactly where that line falls explains most of the frustration people have with ChatGPT's memory.

What crosses the gap

Saved memories. Roughly 8–12 short entries visible under Settings → Personalization → Memory. These are injected into every new conversation.
Custom instructions. The two free-text boxes you filled in once and probably forgot about. These apply everywhere and are, quietly, the most reliable memory ChatGPT has.

What does not

The conversation itself. Every message in a thread is working context and dies with the thread.
Files you uploaded. Scoped to where you uploaded them.
Reasoning it did. Conclusions reached in one thread do not become facts in the next.
Tone you spent an hour tuning. This is the one people mourn.

Why it feels like it forgot you

Because it did not forget. It never stored it.

That distinction matters. Forgetting implies something was held and then lost. What actually happens is that the vast majority of what you say is never written down at all — it is read, used, and discarded when the tab closes. The model is not being careless. It is working exactly as designed, on a memory budget that has to be multiplied across hundreds of millions of users.

The gap between "it has been talking to me for a year" and "it holds eight sentences about me" is where the disappointment lives. More on the numbers in how much ChatGPT actually remembers.

Short answer

ChatGPT remembers a paragraph-length sketch of you across chats, and nothing else. New conversation, same sketch, blank everything else. If you want continuity that survives the thread — and the next model rollout — it has to live somewhere other than inside ChatGPT.

Three ways to widen the gap

Curate the eight slots. Open the memory panel and delete the noise. Every entry that does not matter is one that is crowding out one that does. Free, immediate, and capped.
Use custom instructions properly. They are more durable than saved memories and most people leave them nearly empty. See custom instructions vs memory.
Keep the long version outside. Hold the full picture in a vault you own, then connect it — over MCP, or by pasting a compact export at the top of a thread. The model stops needing to remember, because it can look you up.

Frequently Asked Questions

Stop re-introducing yourself every Monday.

Konshus keeps the long version of you outside ChatGPT and hands it back on demand — so a new thread starts where the last one ended.

Meet Konshus