Relax. We're not going to explain this chapter like a textbook — just like a friend talking you through it, in plain, easy-to-follow terms.
Let's think about it this way for a second
In a long chat, the model tends to forget the rules it started with. That's not because it's careless — it's because the plate is full. It's on you to trim the history. What's worth keeping: the user's latest question, the short rules you set, and 1–2 recent decisions. What's worth dropping: failed attempts, repeated "okay, got it," and long pasted text. Wash off the old dishes on the table before you serve the next course.
Let's connect it to everyday life
Bad — sending the entire message list as JSON. Good — "Rule: say so if you can't find it / So far: explained photosynthesis / Now asking: why does water matter." Sometimes the "New chat" button in the UI is actually a useful tool. If things get too messy, open a fresh chat and re-add the short rules plus short notes. That's not starting over — it's just clearing the table.
Let's try it out together
Keep:
- Rule: if missing, say မတွေ့
- Last fact: explained photosynthesis
- Now: Why do plants need water?
Drop: 40 messages of trial and errorTrim a long chat down to a few key lines.5-Minute Try
Write 12 lines of a fake chat. Keep only 4 for the next question, and explain why you cut the rest.
One Quick Heads-Up
Resend an old mistake, and the model will confidently repeat it.