You gave it your prices on Monday. On Thursday it makes up new ones. You did nothing wrong. The tool has two kinds of forgetting, and each needs its own fix.
It forgets inside a long chat
A chat model can only read a fixed amount of text at once. That amount is called the context window. Every message you send uses some of it, and so does every reply.
When a chat grows past the window, the oldest messages drop out. You can still scroll up and read them. The model cannot. That is why it follows your instructions for an hour and then stops.
Long chats have a second problem. Details in the middle of a long conversation get used less reliably than details at the start or the end. A rule you gave forty messages ago is easy for the model to miss.
It forgets between chats
A new chat starts with an empty page. Unless something was saved, it knows nothing from last week.
ChatGPT has a memory feature for this. OpenAI describes two parts: memories it saved about you, and what it picks up from your past chats. You can see and edit what it holds in the settings.
The catch is that the tool chooses what to keep. Memory is built to know you as a person: your name, your job, how you like answers. Nobody promised it would hold your full price list, your refund policy, or the words you never use.
How to tell which one you hit
- It forgot something from earlier in this same chat. The chat is too long.
- It forgot something from another day. It was never saved to memory.
- It knows the fact but ignores it. The instruction is buried in a long prompt.
What to do about a long chat
- Start a new chat for each task. One endless thread for everything is the main cause.
- Before you leave a long chat, ask for a summary of the decisions. Paste it into the next one.
- Repeat the rule that matters right before the request it applies to.
What to do about facts that must survive
Do not trust memory with facts that cost you money when they are wrong. Keep them in a file you own.
Write one page. Say what you sell, your prices, who you sell to, how you write and what you never say. Paste it at the top of a chat, or put it in the tool's custom instructions. There is a full guide to writing that page.
Then open the memory screen once a month and delete what is wrong. An old fact does more harm than a missing one, because the model uses it with full confidence.
Why this hits one-person businesses hardest
A team spreads what it knows across people. In a business of one, you are the only place it lives. Every tool you open starts from zero, and you are the one who fills it in again.
You are not alone in this. One analysis of 500 Reddit complaints about AI tools put forgetting between conversations at the top of the list. The author sells a memory tool, so read the exact numbers with care.
Nive is being built for this problem: one place that holds how you run your business, so every job starts from the same facts. Join the waitlist to get the demo first.