Why Your AI Forgets or Stops: Claude Artifact Token Limits Explained

What a token is, in plain words

AI models don't read words. They read tokens: small chunks of text, often a whole short word or a piece of a longer one. Every limit on this page is counted in tokens, not words or messages. As a rough rule of thumb, symbols, tags and indents often become tokens of their own, so a screen of code can cost more than a screen of plain English. That's one reason a chat full of code fills up faster than it looks.

The context window: your AI's working memory for one chat

The context window is everything the AI can see while it writes its next reply: your messages, anything you pasted and every reply it has already given. Anthropic's own documentation calls it the model's working memory, separate from what it learned in training. It has a fixed size. When a chat outgrows it, something has to give.

Here's the part most beginners miss. The AI doesn't carry your chat around between messages. Each time you send something, the whole conversation so far goes back in, from the top, and the reply is written from that. A chat with 40 back-and-forths is read in full again on message 41.

Claude artifacts show this clearly. An artifact is the separate panel where Claude puts a file it writes for you, such as a web page or a script. It's still part of Claude's reply, so it counts towards the context window like any other message, and every change Claude writes into the chat adds to what's there. A long web page revised six times is a lot of text for the AI to carry, and it's read again on every new message. That's why an artifact-heavy build hits trouble sooner than a chat of plain questions.

How big is the window? It depends on the model, and it changes whenever new models ship, so check your provider's current documentation rather than a number on a blog page. What stays true: it's finite, and everything counts.

Why your AI forgets long before the chat is full

A long chat goes soft before it hits any hard limit. The more text sits in the context window, the worse the model gets at recalling any single piece of it. Anthropic's engineering team calls this context rot (the same slow loss of memory this page calls long-chat drift), and Anthropic's documentation also says more context isn't automatically better (both checked September 2026). You'll recognise it by these signs:

• It ignores a rule you set at the start ("don't touch the header").
• It brings back a bug it fixed an hour ago.
• It contradicts a decision you both agreed on earlier.
• It rewrites parts of the build you asked it to leave alone.

None of this means the AI is broken or that you prompted badly. Your early instructions are still in the window. They're competing with everything that came after them.

What happens when a chat does hit the wall

When a conversation fills its context window, you'll usually see one of two things. The chat stops: the reply trails off, or you get a message that the conversation is too long. Or the chat carries on, because the tool summarises or drops the oldest messages to make room. Some chat apps, Claude among them, summarise automatically once a chat gets long, on paid plans with code execution switched on in your settings, and Anthropic's developer documentation notes that chat apps can also drop the oldest messages first (both checked September 2026).

When the oldest messages are dropped, their detail is gone. When the chat is summarised, the AI works mostly from a condensed version of the early part, even though Anthropic says Claude still keeps the full history to refer back to. Both keep a long build going. Neither cures context rot, which is why a long build still drifts from what you first asked for. Automatic summarising also costs more: Anthropic's help centre notes that longer conversations that trigger this automatic summarising use more of your allowance.

One special case: if the very first message in a brand-new chat already fails, the most likely cause is that you pasted too much at once, such as a whole project's code or a huge error log. There's nothing earlier to summarise. Paste the part that matters.

The usage limit is a different thing entirely

The usage limit caps how much you can use the AI across all your conversations, not how long one conversation can be. Anthropic's help centre draws exactly this line between usage limits and length limits. If you hit the usage limit, a new chat won't help, because every chat draws from the same allowance. You wait for it to reset, or get more allowance on your account.

Some providers share one allowance across their web, desktop and coding apps (Anthropic does, for claude.ai, Claude Desktop and Claude Code), so a heavy session in one eats into the others. How much you get and how it resets depend on your plan and change often, so read the provider's own current usage page. Look for three things there: whether the limit resets on a rolling window or at a fixed time, whether it's shared across apps, and what happens when you run out.

The two limits are linked in one way. A longer chat uses more of your allowance with every message, because the whole conversation goes back in each time, and Anthropic lists conversation length among the things that drive usage. Starting fresh sooner saves both.

How to tell what went wrong

Work from what actually happened, not how it felt. A cut-off reply, a length message, a forgetful AI and a usage-limit message all feel like "it stopped working", yet each has its own cause and its own fix. Five quick checks, in this order, tell them apart:

1. Read what actually happened. Note the exact message, if there is one, and what the AI did just before it.
2. The reply trails off, or you see a length message: the context window is full. Start a new chat and carry a handover summary forward. If the old chat still replies, ask it for one first (the prompt is below).
3. It forgets instructions or brings back fixed bugs: that's long-chat drift. Ask for a handover summary and start a fresh chat with it.
4. You see a usage-limit message, even in a short chat: it's the usage limit. A new chat won't help. Wait for the reset, or check your plan on the provider's own page.
5. The very first message of a new chat fails: most likely you pasted too much. Trim it to the part that matters and send that.

The flowchart below shows the same causes as the checks above, with the fix for each branch.

Which one did you actually hit?

What to tell your AI: the handover prompt

Two of the four causes in the flowchart have the same fix: move to a fresh chat without losing what you built. Don't retype the project from memory. Ask the AI to write the handover for you, ideally before things go wrong, then paste it as the first message of the new chat.

Before we continue, write a handover summary I can paste into a brand-new chat. Include:

1. What we're building, in two or three sentences.
2. What's finished and working.
3. What's still broken or unfinished.
4. Every rule and decision I've given you (tech choices, things not to change, style rules).
5. The main files and what each one does.
6. The exact next step.

Keep it short enough to read in one minute. Don't include full code; I'll paste any file you need.

What to check in its answer

• Every rule you remember setting is there. If one is missing, that's likely the one it had already lost. Add it yourself before you paste.
• The "still broken" list isn't empty on a build you know has problems. If it is empty, correct it.
• The next step matches what you actually want to do next.
• It's short. A handover as long as the chat defeats the point.

Then open the new chat, paste the summary, and paste only the file you're working on next.

Habits that keep a long build out of trouble

• One job per chat. When you move to a new feature, start fresh.
• Paste the part, not the whole project. The file or section under discussion is enough.
• Start fresh at a milestone, not when it breaks. Once a feature works, ask for the handover and move on while the AI still remembers everything.
• Keep a short project brief you paste at the start of every new chat. Our guide to giving your AI project context shows what goes in one.

What isn't possible

• You can't keep one chat sharp forever. Every window is finite, so a good handover summary is the best carry-over there is.
• A bigger window doesn't cure the forgetting. More room isn't better attention.
• A new chat doesn't reset a usage limit. Only time, or more allowance on your account, gets you going again.