Why ChatGPT forgets everything (and the system that doesn't)
It ignored the instructions you gave it an hour ago. It has no idea who you are in a new chat. And sometimes its saved memories just vanish. These are three different problems. Here is what causes each, what helps, and what to stop expecting.
Everyone who uses ChatGPT seriously eventually has the moment. You spent an hour building something together — a plan, a codebase, a piece of writing — and then it contradicts a decision you made twenty messages ago, as if the decision never happened. Or you open a new chat and the collaborator who knew everything about your project greets you like a stranger.
The frustration is real, but it gets misdiagnosed constantly, because "ChatGPT forgot" describes three different failures with three different causes. Once you can tell them apart, you know what actually helps — and what is snake oil.
The three kinds of forgetting
- Within a conversation: ChatGPT loses track of things said earlier in the same thread. Cause: the context window.
- Between conversations: each new chat starts mostly from zero. Cause: that is the default design, and the memory features only partially soften it.
- Saved memories vanishing: the details ChatGPT stored about you disappear without warning. Cause: unclear, but documented user reports say it happens.
Let us take them one at a time, in plain language.
1. Forgetting within a conversation: the context window
ChatGPT does not have a memory the way you do. What it has is a context window: a fixed-size working area that holds the conversation it can currently "see." Every message you send and every reply it writes goes into that window. The window's size is measured in tokens (chunks of words) and varies by plan and model, but the crucial fact is the same everywhere: it is finite.
When a long conversation outgrows the window, something has to give. The oldest parts of the thread quietly fall out of view, or get squeezed into a rough summary. There is no warning banner. The model just stops seeing your early messages — which is why it suddenly reinvents things you settled long ago, or "forgets" the persona and constraints you set up at the start.
It is a bit like working at a desk of fixed size. New pages keep arriving, and when the desk is full, the oldest pages slide off the far edge. The model is not being careless. The pages are literally no longer on the desk.
2. Forgetting between conversations: amnesia by default
The second kind feels more personal: you have hundreds of conversations behind you, and the new chat knows none of it. This is not a malfunction. By default, every conversation starts fresh, with nothing carried over.
OpenAI has layered memory features on top of that default, described in its Memory FAQ. Two things can carry across chats:
- Saved memories: short distilled facts ("is vegetarian," "works in marketing"), capped at roughly a sticky note's worth of text. I wrote a whole piece about what happens when that fills up.
- Chat history reference: when enabled, ChatGPT can pull relevant details from your past conversations into a new one. Useful — but OpenAI itself is explicit that it does not retain every detail from past chats. It selects what seems relevant, and it may not select the thing you cared about.
So even with everything switched on, what crosses between conversations is a thin distillation: highlights chosen by the system, not the actual substance of the work. The three-thread plan you built last month is not "remembered" anywhere — at best, a couple of sentences about it might surface if the system judges them relevant.
3. When saved memories vanish
The third kind is the one that breaks trust. Users have repeatedly reported opening ChatGPT to find their saved memories simply gone — the memory panel blank, months of accumulated context erased, with no warning and no undo. TechRadar covered one such wave, sparked by Reddit threads where user after user replied "same here" — across free and paid accounts, on desktop, web, and mobile.
I am not claiming this happens to everyone, or often. Most users never see it. But it has happened in documented waves, there is no user-facing backup or restore for the memory feature on most plans, and when it happens there is usually nothing support can do. That combination is what matters: the memory is convenient, and it is also entirely outside your control.
The system that doesn't forget: a memory layer outside the apps
Notice the pattern across all three kinds of forgetting. In every case, the fix that actually works is the same move: keep the important context somewhere you control, outside the AI app. Summaries you carry between threads. Project notes you paste in. Backups of the memory list. You end up building a memory layer by hand.
That is the thing I eventually got tired of doing manually, and it is why I built ThreadRecall — a menu bar app for Apple Silicon Macs (macOS 13 or later, currently a free beta) that does that job automatically:
- It captures the full conversations. As you work in the ChatGPT, Claude, Gemini, Codex, and Perplexity desktop apps, conversations are saved into a local SQLite database on your Mac. Not summaries the system chose to keep — the actual threads.
- Nothing leaves your machine. The archive is local. There is no server-side memory to be wiped in some mysterious wave.
- You can find things again. Keyword search plus optional semantic search, so a half-remembered idea is enough to pull up the thread.
- You can bring context back. Inside Claude,
trecall:commands recall past conversations into the current one — which is re-priming, minus the manual copy-paste. - Switching apps stops costing you the thread. Because capture works across ChatGPT, Claude, Gemini, Codex, and Perplexity, the context is not trapped inside whichever app you happened to use that day. Start in ChatGPT, continue with the same background in Claude.
- Obsidian, if you want it. Conversations can sync into your vault as structured Markdown notes.
To be exact about the claim: ThreadRecall does not change how ChatGPT's context window or memory works — nothing can do that from the outside, and you should be suspicious of anything that says it can. The apps will keep forgetting. The difference is that once your conversations live in an archive you own, their forgetting stops being your loss.
ThreadRecall keeps a local, searchable copy of your ChatGPT, Claude, Gemini, Codex, and Perplexity conversations on your Mac — so the thread survives even when the app forgets. Free beta, early access open.
Frequently asked questions
Why does ChatGPT forget things from earlier in the same conversation?
Every conversation runs inside a fixed-size working area called the context window. When a long thread outgrows it, the oldest messages silently fall out of the model's view. There is no warning when it happens.
Why doesn't ChatGPT remember my previous conversations?
Each chat starts fresh by default. With memory features enabled, saved memories and selected details from past chats can carry over — but those are distilled highlights, not the conversations themselves. It never rereads your full history.
Can ChatGPT's saved memories disappear on their own?
There are documented waves of users reporting saved memories vanishing without warning and without a recovery option. It is not the norm, but it happens often enough that backing up anything important is the safe assumption.
How do I stop ChatGPT from forgetting?
You cannot fully stop it; forgetting is built into how these systems work. You can reduce the damage: keep threads focused, re-prime new chats with a short summary, and keep copies of important conversations outside the app.
Does ThreadRecall stop ChatGPT from forgetting?
No — no tool can change how ChatGPT's context window or memory works, and I would not trust one that claims to. ThreadRecall works from the outside: it keeps full local copies of your conversations across five AI apps, so when an app forgets, you have not lost anything.
