# Push context out on purpose

> **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.

## What you need

```
□ 30 minutes
□ One conversation window (start a fresh one)
□ Some long text to paste — anything will do
   · an article, a manual, meeting notes
   · nothing confidential
```

---

## Step 1 — Plant a marker (5 min)

At the **very start** of a fresh conversation, plant something you will ask for
later.

```
You 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.
```

**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.**

---

## Step 2 — Fill it up (10 min)

Now make the conversation long. Paste your text in **several chunks** and ask for
a tidy-up each time.

```
Tidy the following into meeting-note form.

(paste a long chunk)
```

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.

### What to look for

```
□ Are the answers getting shorter?
□ Is the initials rule starting to wobble?   <- usually the first to go
□ Is it getting slower?
```

---

## Step 3 — Ask for the marker (5 min)

Now check what you planted. **Do not restate the rules.** Just ask.

```
What was this project's codename again?
And what was our rule for writing people's names?
```

### What to look for

| 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](/guide/ai-context-overflow). The model does not say it forgot. It fills the
gap.

---

## Step 4 — Bring it back (10 min)

Now test the remedies one at a time and see **which actually works.**

### Method A — Just ask again

```
Do you remember the rules from earlier?
```

Usually not. **Asking for something that is gone does not create it.**

### Method B — Restate the constraint

```
[Rules, again]
- people as initials only (John Smith -> J.S.)
- the project codename is "Blue Harbour"

Using those rules, tidy the following.
(content)
```

**This works.** It is why long jobs restate their important constraints
periodically. It looks like busywork; it is the standard practice.

### Method C — Summarise and start fresh

```
Pull everything we have tidied so far into one document.
Then list the rules we have been following.
```

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** |

---

## Self-check

```
□ 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
```

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.

---

## Check yourself

**1. What is the dangerous symptom when context is pushed out?**

<details>
<summary>Answer</summary>

**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.
</details>

**2. Why does "do you remember?" not work?**

<details>
<summary>Answer</summary>

**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.
</details>

**3. Why is periodic restatement standard on long jobs?**

<details>
<summary>Answer</summary>

**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.
</details>

---

Next, what to keep beyond the window →
[Memory that outlives the session](/guide/ai-memory)