When should you start a new AI chat, and when should you continue?

A small, practical question that hides a big design decision. Every long conversation with an AI assistant is a balance between context (helpful) and clutter (degrading). This article gives a working rule of thumb.

The hidden cost of a long chat

Every message you send is processed against the entire prior conversation. That’s a feature — it’s what lets the model remember your style, your project, your earlier instructions. It’s also a tax: the more tokens behind the scrollbar, the more attention the model has to spread across them, and the more likely it is to surface a stale detail at the wrong moment.

[Migration in progress — full article body to be brought across from the original Notion source.]

A working rule

• Stay when the new turn depends on what was said three or four exchanges ago.
• Reset when the new turn is a different task wearing the same hat.
• Branch (a new chat seeded with a one-paragraph summary) when you want to keep momentum without the baggage.