There is no error message. You just watch it happen: Gemini forgets a decision you made twenty messages ago, in the same chat, on a plan advertising a context window of up to one million tokens.

Before you blame the model, rule out the two causes that have nothing to do with context length. In a surprising number of cases the conversation is fine and you are simply somewhere that never had memory to begin with.

Quick answer: Check the surface first. Temporary Chats, Gems and Live do not use memory of past chats by design. Then check that Personal context is turned on, since with it off every chat starts cold. If both are fine and you are deep in one long thread, move the work to Google AI Studio, where the token count is visible.

First, work out which kind of forgetting this is

Three different things get reported as "Gemini forgot." They look identical from your side of the screen and have completely different fixes.

	You are on a memory-free surface	Personal context is off	The thread outgrew its budget
What you see	It knows nothing about you or past chats	Every new chat starts from zero	It forgets things from earlier in this same chat
Where it happens	Temporary Chats, Gems, Live	Everywhere	Any long thread
Is it by design?	Yes	Yes, it is a setting	Yes, it is the context window
Does a new chat help?	Only if you leave the surface	No	Yes
The fix	Move to a normal chat	Turn Personal context on	Split the work, or use AI Studio

The first column is the one almost nobody checks, and it is the most commonly missed cause. Temporary Chats, Gems and Live chats all skip memory of past chats deliberately. Temporary Chats go further: they do not appear in your recent chats or in Gemini Apps Activity, they are not used to personalize Gemini, and they are kept for up to 72 hours. Everything works exactly as specified, which is precisely why it is confusing.

Fix it: rule out the cheap causes, then manage the thread

Work down this list in order. The first three take under a minute combined and resolve the majority of cases.

1. Confirm you are not in a Temporary Chat. If the conversation is missing from your recent chats list, that is your answer. Temporary Chats are isolated on purpose. Start a normal chat and re-establish what matters.
2. Confirm you are not in a Gem or a Live chat. Gems are excellent for a fixed job with fixed instructions, and Live is built for talking. Neither carries memory of your past chats, so context you established in a regular conversation will not be there.
3. Turn Personal context on. With it enabled, Gemini references past chats to learn your preferences and carries them into new conversations. It is available in the Gemini mobile app, the web app, and Gemini in Chrome where offered. With it off, cross-chat memory simply does not exist for your account.
4. Ask explicitly instead of assuming. There is a well-documented behavior where Gemini appears to forget something, you type "recap," and it suddenly remembers. That is not the model recovering its memory, it is the app performing a broader retrieval because you asked it to. Treat it as a tool: when something matters, name it rather than hoping it carried.
5. Restate constraints close to your question. In a long thread, instructions from message three compete with everything after them. Putting the rule that matters at the end of your current message gives it the best chance of being honored.
6. Split long work into focused threads. One thread per task beats one enormous thread across five tasks. Before you split, ask Gemini for a handoff note covering the goal, the decisions made and the next step, then paste that into the new chat.
7. Move document-heavy or genuinely long work to Google AI Studio. aistudio.google.com gives you direct access with explicit system prompt controls and full token count visibility that the consumer app does not expose. If you need to know how much room you have left, this is the only place you can see it.

None of this requires a paid plan or an extension, and steps one through three cost nothing but attention.

About that one million token window

This is the part of the story that deserves care, because it is easy to overstate.

Google advertises a context window of up to 1 million tokens on Gemini Pro and Ultra plans, describing it as up to 1,500 pages of text or 30,000 lines of code. That is the model's stated capability.

What users report about the app is different. A thread on Google's own developer forum describes memory coherence breaking down somewhere around 150k to 200k tokens in the Gemini app, while the same user pushed past 500k tokens in Google AI Studio with the model still quoting phrases from early messages accurately. Android Authority covered the same discrepancy in June 2026, reporting user claims that the active conversational context in chat is bottlenecked well below the advertised window, and noting that Google was asked about the gap.

Here is the honest framing, and we are going to be strict about it. These are user reports and press coverage of user reports. Google has not publicly confirmed a lower in-app limit, has not reconciled the difference, and no specific in-app number should be treated as fact. What you can reasonably conclude is narrower and still useful: the advertised figure describes what the model can do, not a promise about how much of your chat history the consumer app keeps active. If your work depends on knowing where the edge is, AI Studio is where you can actually see it.

Why a thread degrades before it breaks

Gemini rarely gives you a hard error the way ChatGPT or Claude do. It degrades quietly instead, which is worse in one specific way: there is nothing to search for when it happens.

Two mechanisms are at work. Once a single thread runs past the token budget, the oldest turns stop influencing new answers even though the messages still display on screen. The app stores the text and shows it to you, and that visibility is exactly what makes the failure feel like a betrayal.

Before that point, context rot is already setting in. As a conversation grows, attention spreads across more tokens and older material starts competing with the thing you actually asked about. Quality drops well before any ceiling is reached, which is why a bigger window is not automatically a better experience. Compaction, the summarize-and-continue mechanism that several assistants use, buys room at the cost of detail.

The full mechanics, across every assistant, are in why AI assistants forget your conversation.

Stop having this problem

The checklist above genuinely fixes the surface and settings causes, and thread hygiene handles most of the rest. But notice what you are doing: auditing which mode you are in, remembering to say "recap," restating constraints, writing handoff notes, watching a budget the app will not show you.

The structural problem is that your context lives inside a thread, inside one app. The thread is the unit of memory, so when it ends the memory ends with it. Personal context helps, and it is still a layer bolted to the side of that same architecture.

folk is built the other way around. It compacts long conversations, which is not unusual, but the useful difference is that its memory is not scoped to a thread at all. It reaches you in iMessage and Telegram, so there is no conversation to reach the end of and nothing to re-paste when one dies. What you told it last month is still there. The common alternative, a browser extension that adds a memory layer to someone else's chat window, does nothing the moment you pick up your phone.

If you have context worth keeping in another assistant, you do not need to retype it. You can import your ChatGPT and Claude memory in one paste.

Hitting this somewhere else? See the guides for ChatGPT and Claude, or the overview of why every AI assistant forgets.