Why your AI forgets the rules by the fortieth message

You set the rules at the start. Use this year's price list. Never promise a delivery date. Write in British English. For the first ten messages the assistant behaves. By message forty it quotes last year's price, promises Friday, and spells colour without the u.

Nobody changed the instructions. They are still sitting at the top of the chat. The model has simply stopped giving them much weight.

This is measured, and it is not one bad model

In May 2025 a team from Microsoft Research and Salesforce published a paper called "LLMs Get Lost in Multi-Turn Conversation". They took tasks that leading models handle well in a single prompt and fed the same information in over several turns, the way people actually talk to a chatbot. Average performance fell by 39%. Unreliability, the spread between a model's best and worst runs, more than doubled. Every model they tested showed the drop. Their own phrasing was that when models take a wrong turn early, they get lost and do not recover.

Chroma's "Context Rot" report (July 2025) covers the other half. As the input gets longer, models recall what is inside it less reliably, even on simple lookups. A long chat is a long input, and your rules sit at the far end of it.

Put those two findings next to each other and the fortieth-message problem looks less like bad luck and more like the expected result of keeping rules inside a conversation.

Why "just remind it" does not hold

The usual fix is to paste the rules again when things slip. It works for a while. It also depends on one person noticing the slip, remembering the whole rule set and pasting it into one chat. The colleague working in the next chat has her own version of the rules, typed from memory, missing two of them.

There is a second problem. Because the rules live in the conversation, they age with it. The model's own earlier answers, including the wrong ones, sit closer to the next reply than your instructions do. A wrong answer at message twelve becomes context for message thirteen.

Move the rules out of the chat

The pattern that holds up is to keep rules and facts in files that sit outside any conversation and have every session read them before it starts. Anthropic's engineering team described the same approach in "Effective context engineering for AI agents" (September 2025): keep structured notes outside the context window and bring them back in when they are needed, rather than trusting a long transcript to carry them.

This is how we run our own AI work at Provecta Group. Company rules, approved facts and past corrections sit in a written memory that the company owns. Every working session loads it before it does anything. When a session drifts, we close it and start a fresh one, and the fresh session loads the same rules from the same place. A correction written into that memory once reaches every later session, whoever is running it.

A few habits make it work:

• Rules are short, written as instructions, and each one carries its reason. "Quote only from the current price sheet, because the old one undercharges" holds up better than "be accurate".
• Facts carry a date and an owner, so a stale one can be spotted and someone is answerable for fixing it.
• Long jobs are split into steps, and each step starts by reading the rules again instead of scrolling back through the chat.

Where written rules stop helping

Written memory makes drift less likely. It does not make a model obey. A model can read a rule and still break it, especially when the task pulls hard the other way. For anything that costs money or reaches a customer you also need a check that can stop the output: a person approving it, a script that refuses any price not on the sheet, or a permission the AI simply does not have. Memory tells the AI what the rules are. A gate decides whether its work goes anywhere.

It also will not rescue bad rules. If the price sheet in your memory is wrong, every session will now be wrong in the same way, and more confidently than before.

One thing to try this week

Pick the AI task your team runs most often. Write its rules on a single page: what it must always do, what it must never do, and why, one line each. Save that page where everyone can open it. Start every session for that task by loading the page into a fresh chat instead of continuing an old one.

Then watch where it still slips. Those slips are the rules that need a hard check rather than a reminder.

To see where your team's AI use is most exposed, try the AI readiness check, or book 20 minutes at pgco.world/book.