One long AI chat per ad account, or a fresh one with a brief each time?

I've been going back and forth on how to set up the chat itself when you run paid search through an AI assistant. Some people keep one long thread per account so the model remembers past changes, naming conventions and what got tested. Others start fresh each session and paste in a short brief. My hunch is the long thread works well for analysis and badly for changes. After a few weeks the context is full of old decisions that no longer apply, and the model can treat a paused test as current. A fresh chat with a one-page brief (goals, target CPA, what's off limits, last major change) feels safer for anything that writes to the account. Does anyone have a rule for when to reset, or is telling it to ignore old context good enough?

Agreed on the fresh chat for anything that writes. Telling the model to ignore old context doesn't really work, because those messages are still in what it reads, and it will often treat a paused test as live. Say a plumbing company ran a call-only ad test in spring and paused it when cost per lead climbed. Six weeks later the old thread still says "test running," and a suggestion built on that is easy to get.

I'd add dates and status to the brief, not just the goal. Something like "Call-only test paused 14 March, cost per lead over $60, don't restart without fresh data." Then have the model pull the current campaign status before it proposes any change, and treat that live state as the source of truth over both the brief and the old thread.

Mostly agree, but it hinges on whether the assistant can read the account live. If it can, your approach works: the status pull settles whether the test is running, and the brief only has to explain intent. If it can't, the model only knows what you paste in, and a dated brief is only as current as the last time someone updated it. Then it's guessing about that paused test no matter how carefully you word the brief.

In that case I'd keep a changelog outside the chat, one line per change with the date and the reason it stopped or started. For each fresh session, paste in the last few entries instead of the whole history, so the model gets the decisions that still bind and nothing else. The reset habit matters less than whether that log exists and actually gets updated.

The log matters more than the reset rule, agreed. I'd add one field to each entry: the condition that would reverse the decision. Something like "Paused call-only test, cost per lead over $60" lets the model check whether that condition still holds before it treats the pause as permanent. A date and a reason only tell it what happened, not whether it still applies.

Once entries carry their reversal conditions, you can prune the log too. Keep the rows whose conditions are still live and drop the rest. Which do people skip more often when they write these, the reason or the condition?