What you will learn Not a chapter to read but one to do. You will deliberately create the moment when the front of a long conversation quietly falls away, see which symptoms appear, and find out what actually brings it back.
At the very start of a fresh conversation, plant something you will ask for later.
Planting two rules is the device of this exercise. One is a format (initials) and one is a fact (the codename). You get to watch which of the two collapses first.
Now make the conversation long. Paste your text in several chunks and ask for a tidy-up each time.
Repeat this five to eight times, a different passage each round.
This step is tedious, and that is the point. Context fills up tediously in real work too — it does not fill up all at once.
Now check what you planted. Do not restate the rules. Just ask.
| The answer you get | What it means |
|---|---|
| Both, exactly right | still inside the window. Do more of step 2 |
| Codename right, initials rule vague | it has started to slide |
| "No codename was specified" | the front is gone |
| A confident but different codename | the dangerous failure |
If you get the last one, keep that screen. It is the clearest possible picture of what "quietly disappears" means in when context gets pushed out. The model does not say it forgot. It fills the gap.
Now test the remedies one at a time and see which actually works.
Usually not. Asking for something that is gone does not create it.
This works. It is why long jobs restate their important constraints periodically. It looks like busywork; it is the standard practice.
Paste the result as the first message of a new conversation. The conversation gets short and the rules come back to life.
| Method | When |
|---|---|
| B, restate the constraint | continuing the job you are on |
| C, summarise then start fresh | at the end of a stretch of work |
| A, ask again | do not |
If the third is unchecked, repeat step 2 more and try again. If nothing ever breaks, that model's window is wide — which is also a result. Look at its cost and speed instead.
1. What is the dangerous symptom when context is pushed out?
Inventing something plausible rather than saying it forgot. "I don't know" gets caught by a human; a confidently stated wrong codename goes straight into the document.
2. Why does "do you remember?" not work?
What left the window does not exist. Asking does not restore it — you have to write it back in. That is why the remedy is restating the constraint, not asking about it.
3. Why is periodic restatement standard on long jobs?
You cannot know in advance when it will slide. By the time you notice, a few wrong outputs already exist. Repeating the constraint is far cheaper than that.
Next, what to keep beyond the window → Memory that outlives the session
HyperTeams□ 30 minutes□ One conversation window (start a fresh one)□ Some long text to paste — anything will do · an article, a manual, meeting notes · nothing confidentialYou are going to help me tidy up meeting notes. Two rules.1. Always write people's names as initials only (John Smith -> J.S.)2. This project's codename is "Blue Harbour". Whenever you refer to the project, use that codename.Reply with just "understood" if that is clear.Tidy the following into meeting-note form.(paste a long chunk)□ Are the answers getting shorter?□ Is the initials rule starting to wobble? <- usually the first to go□ Is it getting slower?What was this project's codename again?And what was our rule for writing people's names?Do you remember the rules from earlier?[Rules, again]- people as initials only (John Smith -> J.S.)- the project codename is "Blue Harbour"Using those rules, tidy the following.(content)Pull everything we have tidied so far into one document.Then list the rules we have been following.□ You saw one of the rules genuinely go vague in step 3□ You saw which collapses first, the format rule or the fact□ You confirmed method B works□ You confirmed method A does not <- confirming this is what stops you doing it