You start a chat. It is sharp, it understands the job, you get somewhere. 2 hours later the same conversation is producing mush, contradicting decisions you made at the top, and asking about things you already told it.
Nothing broke. You hit a limit that nobody explains, and there is a 10-second fix.
The confusion comes from 3 different things that all sound like memory and behave completely differently. Once you can name which one failed, you know what to do about it.
The 3 things
Context is everything the AI can see at the moment it answers you: your message, the conversation so far, files you attached, your saved instructions. If it is not in the context, it does not exist, no matter how obvious it feels to you.
The context window is how much context it can hold at once. Large, but not infinite. In a very long conversation the earliest messages fall out the back to make room, which is why hour-3 answers start ignoring hour-one decisions.
Memory is different again: a longer-term record the AI keeps about you across separate conversations. It builds itself as you chat. Claude has this on by default and you can view, pause or clear it in settings. It holds the durable stuff, what your business is, how you like things written. Context holds the live stuff, what you are working on right now.
The clean way to hold it: memory is what it knows about you. Context is what it is looking at.
Which one failed you
| What you noticed | The culprit | What to do |
|---|---|---|
| New chat treats you like a stranger | Instructions not set, or memory has not built yet | Fill in the instructions box. That works immediately |
| Long chat forgot a decision from earlier | Context window. The message fell out the back | Run the handoff prompt below, start fresh |
| It contradicts something you said 10 minutes ago | Context window, usually with a big paste in the middle | Same fix. Also stop pasting things it does not need |
| It does not know a fact about your business | Never had it | Paste the fact, or put it in a project |
| It knows something about you that you did not tell it this chat | Memory doing its job | Fine. If it is wrong, ask it to update memory |
| Answers got shorter and blander over a long session | Context window filling up | Fresh chat |
| It "remembers" something you asked it to forget | Memory, and it needs clearing | Settings, memory, edit or clear |
Only 2 rows on that list are really the tool's limits. The rest are setup.
The handoff prompt
When a long conversation starts going soft, do not fight it. Do not re-explain. Type this:
Summarise this conversation so I can continue it somewhere else. Include: what we are trying to do, every decision we have made and why, anything we explicitly ruled out, the current state of the work, and what is still open. Write it so that pasting it into a fresh chat lets us carry on exactly where we left off. Do not summarise it for a reader, write it for you.
Copy the answer. Start a new chat. Paste it. Carry on with a full window and a sharp model.
That last instruction, "write it for you", matters more than it looks. Ask for a summary and you get something readable and lossy. Ask it to write a handoff for itself and you get the decisions and the ruled-out options, which is the part you actually need.
Habits that stop it happening
One topic per chat. Chats are free and unlimited. The instinct to keep one long running conversation is the single biggest cause of this problem. Start new ones constantly.
Do not paste what it does not need. Every word you paste spends window. Pasting a 40-page document to ask about one clause is the fastest way to a vague conversation. Paste the clause.
Put durable things where they belong. If you find yourself explaining your business again, that belongs in your instructions. If you keep attaching the same price list, that belongs in a project. Context is for today's work.
Run the handoff before you need it. On anything that will run over multiple sessions, ask for the handoff summary at the end of each one and keep it. Starting tomorrow with yesterday's handoff pasted in is a better start than scrolling back through yesterday's chat.
What memory is and is not
Worth being clear, because people either ignore it or worry about it and both are wrong.
It is a running set of notes the AI keeps about you, built from your conversations. It makes chat 20 better than chat one. You can see exactly what it holds, and you can change or delete any of it.
2 practical notes:
If you share an account, decide about memory together. 2 people's preferences in one memory produces an assistant that is slightly wrong for both of you. Either turn it off or keep the notes about the business rather than the person.
Check it occasionally. Ask "what do you currently remember about me and my business?" every month or so. It is a 2-minute read, and the usual finding is something out of date rather than something alarming.
What the handoff loses
The handoff loses nuance. It carries the decisions, not the texture. On creative work where the tone of the conversation mattered, expect the new chat to feel slightly different and be ready to nudge it back.
A summary of a confused conversation is a confused summary. If the chat went off the rails and you never corrected it, the handoff carries the mess forward. Sometimes the right move is to start clean and re-brief properly.
Memory does not travel between platforms. What Claude knows about you, ChatGPT does not, and vice versa. If you run both, either accept that or keep one instructions block you paste into each.
Nothing here fixes a fact it never had. Context problems and knowledge problems feel the same from the outside and have different fixes. If it has never been told your prices, no amount of chat management will help.
Rescue your longest chat
Open your longest running chat, the one that has gone a bit dull, and run the handoff prompt on it. Then start a fresh one and paste the result. The difference in the next answer is usually obvious.