You know the feeling. You started a chat with your AI assistant weeks ago — your planning session, your study notes, your half-written project — and at some point the answers quietly got worse. It repeats things it already told you. It ignores instructions you gave it back on day one. Ask it about something from earlier in the chat and it bluffs confidently or asks you to repeat yourself.
The chatbot didn’t get dumber. It ran out of room. Every AI chatbot can only hold a fixed amount of conversation in its working memory at once, and when your chat exceeds it, the oldest parts of your conversation are the first to go.
A chatbot doesn’t remember — it re-reads
Here’s the mental model that makes everything else click. When you send a message, the chatbot doesn’t “remember” the earlier conversation the way you do. Instead, the app bundles up the whole chat history — your messages, its replies, everything — and hands it to the model as one big block of text along with your new message. The model reads that entire block fresh, every single time, then writes its reply.
That block has a hard size limit. It’s called the context window, and it’s measured in tokens — the small chunks of text the model actually processes. As a rule of thumb from OpenAI’s own documentation, one token is roughly four characters, or about three-quarters of an English word. A long chat of a few thousand words is tens of thousands of tokens, and every message you add makes the pile taller.
The context window is far bigger than it used to be — modern models advertise windows of hundreds of thousands or even over a million tokens. But “bigger” is not “infinite,” and in practice your consumer chatbot may use only a portion of what the model technically supports, especially on free plans. The limit still exists, and it’s still quietly shaping your chats.
What a full context window looks like from your side
Your chatbot almost never announces that it’s out of room. It just adapts, and you notice the symptoms before you ever learn the cause:
It forgets instructions from early in the chat. You told it on page one to write in a casual tone and keep answers short. Three weeks later it’s writing formal essays. The instruction fell out of the window.
It repeats itself or contradicts itself. Earlier answers it could have checked are no longer in front of it, so it regenerates from scratch and the new version doesn’t match.
It mixes up two things you discussed. Details from your budget spreadsheet and your holiday itinerary blur together because only fragments of each are still in the window.
Its answers get generic. Specific, useful answers need specific context. Strip the context away and you get what the model falls back to: smooth, confident generalities.
Sometimes, in a very long chat, the bot stops accepting new messages entirely or starts erroring. OpenAI’s own troubleshooting guidance for ChatGPT lists “start a new chat if the conversation is long or has many turns” as one of the first fixes — which tells you how routine this failure is.
One thing worth clearing up: this is not the same as the chatbot’s saved memory feature. Saved memory (where a bot keeps notes about you between chats) is a separate system — we wrote a full breakdown of how saved chat memory is different from conversation memory. The context window is per-chat: only what happens inside this conversation counts.
Starting fresh is the fix — but don’t start empty
The most effective thing you can do is also the most obvious: start a new chat. A fresh conversation has a full, empty context window, and the answers sharpen up immediately. But “start over” sounds like losing weeks of work, so here’s how to move a long project across without losing anything.
Get the chatbot to summarize itself before you leave. Before abandoning the old chat, ask: “Summarize everything important from this conversation so I can continue it in a new chat — include my goals, the decisions we made, and any constraints.” A good handover summary is a few paragraphs, not a novel. Paste that into the new chat and you’re back up to speed in seconds.
Keep one project per chat. The people who hit context limits most often are the ones running an entire life out of a single thread — the same chat for meal planning, coding help, and travel advice. Every off-topic detour eats space that the main project needs. One thread, one topic, and the window lasts far longer.
Move standing instructions somewhere permanent. If your chat keeps relying on the same preferences — tone, format, your role, your constraints — don’t re-explain them in every thread. Put them in custom instructions (or the equivalent on your chatbot of choice) so they travel with you into every new chat for free, without costing window space.
Don’t paste the whole old chat into the new one. It’s tempting to copy twenty pages of history “just in case.” All you’ve done is fill the new window with old noise. The summary approach works better because it keeps only what still matters.
A couple of habits that delay the problem
You can’t raise the limit, but you can spend it more wisely. Ask for shorter answers when you don’t need essays — long replies consume the window just as fast as long questions. When a chat is about one document or decision, keep follow-ups focused instead of wandering through unrelated topics between turns.
And if you notice answers getting sloppy, don’t just keep pushing. Ask the bot to summarize the important points so far, copy that summary somewhere safe, and start a clean chat. It takes thirty seconds, and it feels like getting a new assistant.
Finally, the same rule from our guide to checking a chatbot’s answers applies double here: the longer the chat, the more you should verify anything important against the source. A chatbot that’s quietly losing context is a chatbot that’s more likely to fill gaps with confident guesses. Catch the drift early, move to a fresh chat, and the quality comes right back.
