You open ChatGPT. You ask it to keep working on the project the two of you were deep in last Tuesday. It replies, in that polite voice it uses when it has no idea what is going on, that it would be happy to help and could you please provide some context.
So you explain the project again. Names, constraints, file structure, preferences. Ten minutes later, when it finally starts being useful, you have already done a chunk of the work the AI was supposed to do for you. Every session opens with this same tax.
If this feels like the entire promise of an AI assistant inverted, you are not imagining it. The memory of an LLM assistant is, by design, shorter and more fragile than the marketing suggests. That is not a bug. It is a structural feature of how these systems work, and it is the single biggest reason your AI workflow keeps stalling.
Quick takeaways: ChatGPT now ships with memory on by default for Free, Plus, Pro and Team users. GPT-5.5 carries a 1 million token context window, but effective attention collapses long before that. Your saved memories live inside OpenAI and do not move to Claude, Gemini, or anything else. The real fix is an external memory layer that any model can read.
The "wait, didn't I tell you that?" moment
Every long-running ChatGPT user has had this experience. You said something specific in chat seven, ChatGPT confidently uses it in chat eight, then in chat fourteen it has no idea what you are talking about. It is not gaslighting you. It genuinely cannot see that context any more.
The reason is mechanical, not philosophical. Every LLM runs inside a context window, a finite buffer of tokens that holds the active conversation. GPT-4o sits at 128,000 tokens, roughly 96,000 words. GPT-5.5, rolled out in the API on April 23, 2026, jumps to about 1 million tokens, somewhere around 750,000 words. Cross the limit and the oldest material drops off the back. The model does not store it. The model does not summarise it. The model simply does not see it any more.
And here is the part the headline number hides. A 1 million token window does not mean a 1 million token memory. Independent retrieval tests on long-context models show a sharp U-shaped attention curve: the model holds the first few thousand and last few thousand tokens, and the middle goes effectively dark. Past around 8,000 tokens of genuine use, recall starts dropping fast. The window is huge. The room with the lights on is not.
This is why long sessions feel like they get dumber over time. They are not getting dumber. They are losing the front of the conversation as the back fills, and quietly losing the middle along the way.
What ChatGPT actually remembers (and what it doesn't)
Be precise about what the system can and cannot do. As of 2026, ChatGPT has three memory layers, each with its own limits.
1. The context window (per session)
Inside one conversation, the model sees everything that fits in the window. For GPT-4o that is 128k tokens. For GPT-5.5 it is up to 1M tokens via the API. Once it is full, the front rolls off, and well before that the middle starts to fade.
2. Saved memories (across sessions)
OpenAI first announced memory in February 2024, then rolled it out broadly on September 5, 2024 to Free, Plus, Team and Enterprise. This is a curated list of facts the model decides are worth keeping ("User is allergic to peanuts", "User prefers Python") and surfaces into later chats. You can edit the list in Settings. Independent testing puts the practical ceiling around 100 to 150 saved memories before older ones get pruned.
3. Reference chat history (across sessions, since April 2025)
This is the implicit layer, launched on April 10, 2025. The model can pull patterns from past conversations even when no specific fact was saved. There is no per-item quota here. There is also no per-item list you can audit.
Stacked together, those three layers are a real upgrade over the stateless ChatGPT of 2023. They are still not enough to solve the problem you actually have.
Why 2026 persistent memory still falls short
Three reasons. They show up in user complaints every week.
- Saved memory has a ceiling. Users keep hitting "saved memory full" around the 100 to 150 mark. After that, new facts either get dropped or quietly displace older ones you cared about.
- The model chooses what to save, not you. If the system decides a fact is not memory-worthy, it does not store it, and you usually only notice two weeks later when it cannot help you.
- Reference chat history is invisible. You cannot ask "please print me the list of what you think you know about me." It is in there. It is not editable item by item. It is sometimes wrong.
Put the three together and you have memory that is real, but not yours. It is OpenAI's interpretation of what should be remembered, surfaced when OpenAI's model thinks it is relevant, in a place you cannot fully see.
Your memory is locked inside OpenAI
Here is the part most people miss until they need it. The saved memories you have built up with ChatGPT live inside OpenAI. They do not travel.
Switch to Claude because Anthropic shipped a model that is better at your task this month? You start from zero. Open Gemini because it is sharper at coding this week? Start from zero. Run a local model for privacy? Same. Every assistant ships with its own memory layer, none of which talk to each other, none of which let you carry the corpus forward.
The best model now changes every two or three months. Your accumulated context is the most valuable thing you have, and it is the worst thing to leave on one vendor's servers.
The five memory limits every AI will hit
Pick any assistant. They all break in the same five places. Once you see the pattern, you stop blaming the model and start designing around it.
- The context window. Every conversation has one. Once it fills, content rolls off, and the middle fades long before the front does.
- The platform lock. Memory built on one model does not move to another.
- The implicit-saving bias. The model picks what to remember; you only sometimes get a vote.
- Fact decay. Tell the model something six months ago, never reinforce it, and it can fade even if it is technically still in the store.
- The quota. Saved memory is finite and fills up faster than people expect.
What actually works in 2026
The pattern that solves all five at once is to stop treating the AI model as your memory and start treating it as a query engine over an external memory you control.
The model is not the memory. The model is a window. Windows do not keep things; rooms do.
Concretely: keep your durable context (documents, photos, voice notes, decisions, contacts, receipts) in a separate, AI-readable memory layer that you own. When you start a session with any model, the model reads from that layer on demand. When you switch from ChatGPT to Claude to a future model, the memory does not move with the vendor. The model changes, the memory stays.
This is the entire reason MemX exists. MemX is not another LLM. It is the external memory layer the LLMs were never going to be on their own. You send documents, voice notes, photos and emails to MemX on WhatsApp. MemX reads them, indexes them, and holds them in your account. When you need an answer ("what medication did my doctor prescribe last week?", "what is the gate number?", "how much was the repair quote?"), you ask MemX in plain English. The corpus is yours. It does not sit inside one assistant's saved-memory ceiling. It does not vanish when you switch models.
In practice this is the boring infrastructure work you already half-do, done deliberately. Documents go to one searchable place. Voice notes go to one searchable place. Receipts, screenshots, business cards, all in the same place. The AI reads from them. The folder structure stays out of it. When the next model arrives, you point it at the same memory and keep working. ChatGPT can still forget you. Your memory does not.
Stop re-explaining your life to ChatGPT every Tuesday. MemX holds the context once, in your account, on the channel you already use (WhatsApp). Free to start at memx.app.
Key takeaway: this problem will not be solved by a bigger context window or a better persistent-memory feature inside one assistant. The problem is that your context lives in someone else's product. Move it into a place you control, and every future model becomes a smarter window over the same memory.
01How big is ChatGPT's context window in 2026?
GPT-4o runs at 128,000 tokens, roughly 96,000 words. GPT-5.5, rolled out in the API on April 23, 2026, goes up to about 1 million tokens. Effective attention is much smaller than the headline number: independent retrieval tests show recall starts dropping past roughly 8,000 tokens of real use.
02When did ChatGPT get persistent memory?
OpenAI first announced memory in February 2024 as a limited test, then rolled it out broadly to Free, Plus, Team and Enterprise on September 5, 2024. Reference chat history, the implicit layer that pulls from past conversations, launched on April 10, 2025.
03Is ChatGPT memory on by default for free users?
Yes. As of 2026, memory is on by default for Free, Plus, Pro and Team. You can turn off saved memories or chat history independently in Settings, Personalization, Memory. Temporary Chat bypasses memory entirely.
04Why can't ChatGPT just remember everything?
Two reasons. Architecturally, every LLM runs inside a finite context window, and compute scales with window size. Economically, persistent storage and retrieval is paid infrastructure, so vendors put quotas on it. The combination is why you hit a ceiling.
05Are Claude and Gemini better at memory than ChatGPT?
Each handles memory differently and each has tradeoffs. The deeper issue is that none of them let you carry your memory between platforms. That is the cross-cutting problem, not the per-vendor one.
06What is the simplest way to keep context across AI assistants?
Maintain your durable context (documents, voice notes, decisions, receipts) in an external, AI-readable place that you own. Reference it from whichever model you are using that week. The model changes; the memory does not.
