Trigger: they've lost something between sessions. Or a long conversation started giving worse answers, repeating itself, or forgetting what was said earlier. Or the opening diagnostic showed the AI's memory of them is patchy or wrong.
Establish context
If they've been having long conversations, ask: have you noticed it getting worse towards the end? Repeating things? Forgetting what you said an hour ago?
If they use it across days, ask: when you open a new chat, how much do you have to re-explain?
They'll almost certainly say yes to one of these.
What changes
A conversation with AI is not like a relationship that builds over time. It's a window. Everything said in this conversation is visible. Everything from a previous conversation is gone, unless something specific carried it over.
Why long chats degrade: every model has a limit on how much text it can hold at once — the context window. Early in a conversation there's plenty of room. As it fills, older parts get compressed or dropped to make space. The model isn't getting tired or bored; it's literally losing access to what was said earlier.
Why memory across chats is unreliable: most providers offer some form of memory — a summary of things learned about you, carried between conversations. This is a compressed note, not a transcript. It's lossy by design: the AI decided what seemed important, wrote a short version, and discarded the rest. It may be out of date. It may be wrong. It may have merged two different things. And you can't see what it chose to keep versus what it dropped, unless you ask.
What to do about it:
- Start a new conversation when the topic changes. Don't run one enormous chat for everything. A fresh conversation with a clear purpose will outperform a long one that's accumulated noise. If you need something from a previous chat, summarise it yourself in the new one rather than hoping it carried over.
- Check what it remembers. Ask it outright: what do you know about me? What it comes back with tells you what survived and what didn't. Correct it — that correction itself becomes part of the memory.
- Don't trust the memory for anything important. If a fact matters — a client's name, a number, a decision you made — keep it somewhere you control. A text file, a note, a spreadsheet. Anywhere that doesn't summarise or discard.
The deeper point: every tool the person relies on for memory — their notes, their files, their CRM — exists because human memory is unreliable. AI memory is unreliable in the same way, for different reasons. The answer is the same: write it down somewhere that doesn't forget.
This connects to → your-real-data.md (giving it your own files) and → repetition-is-a-signal.md (capturing what you keep retyping).