Act 3 · Lesson 10 of 21≈6 minutes○ Saves on this device

The Desk, Not the Brain: Context Windows

By the end you'll be able to

  • Explain what the context window is: the fixed token budget the model can use right now.
  • Explain why an AI "forgets" mid-conversation: material has dropped out of the window, not faded from memory.
  • Manage the window on purpose: summarise, restate, or start fresh.
⚡ 30-second warm-up: from Lesson 9, what gets assembled fresh for every single loop of generation?

Everything: hidden instructions, the conversation so far, anything you've pasted in. The model reassembles that whole bundle and rereads it on every single pass. Today you'll meet the hard limit on how big that bundle can be.

Revisit Lesson 9

Twenty minutes into a chat, the AI forgets your name from message one. Yet it used your name just fine back at minute five. Find out why.

Start with a contradiction

Same assistant, same conversation, no restart. Early on, it used your name correctly three times. Now it asks who you are. It didn't get worse at listening. Something concrete changed about what it could see.

Commit to a prediction

What most likely explains the "forgotten" name?

This lesson changes one thing: you'll stop picturing an AI slowly forgetting like a person. Picture a fixed-size desk instead. Everything it can use right now has to physically fit on that desk, and older pages get pushed off as new ones arrive.

Discover the mechanism

How it works The fixed-size desk

Every inference run (Lesson 9) starts the same way: it assembles one bundle of tokens. That bundle holds the provider's hidden instructions, the conversation so far, and anything you've pasted in. It has a hard maximum size, called the context window. It's the only input the model can actually change on the fly. Everything else is frozen training (Lesson 5).

This has real consequences. A long chat eventually exceeds the budget, so older turns get dropped or compressed. That's the "forgotten name" mystery, solved. A pasted 300-page document might not fit at all, or might fit but get read unevenly: Lesson 8's spotlight doesn't shine equally across a huge desk, so material buried in the middle can be under-weighted. And a brand new conversation starts with an empty desk. By default, nothing carries over.

Nine times out of ten, when people say an AI "remembers" or "forgets", they're describing desk mechanics, not anything like memory.

desk sizes vary a lot between products, and change often. Treat any specific figure as a snapshot, not a rule. Lesson 15 covers the rest of the story: features that genuinely persist information between conversations work differently again, stocking the desk before you even arrive.
Useful mental model An open-book exam at a small desk

Picture a brilliant examinee sitting an open-book exam: that's the trained model. They can bring any materials they like, but only what fits on the desk in front of them is usable right now. Slide a new page on, and something else slides off the far edge.

  • Desk area = the context window, measured in tokens (Lesson 3).
  • Pages on the desk = your messages, its replies, pasted documents, the provider's hidden instructions.
  • Sliding off = older turns being truncated or summarised away.
  • The examinee's trained skill = the frozen weights (Lesson 5): knowledge built from practice, not read off the desk. Training versus context is exactly skill versus papers.

Overflow the desk yourself

This is a deliberately tiny desk (200 tokens), so the limit arrives fast. Feed the chat, watch the gauge, and see what survives.

Desk 0%

Ask it:

Where this bites at work

In the field · Kerry

Kerry's care-plan drafting chat degrades after an hour. The suggestions get vaguer, it repeats a question, it ignores a detail from early on. Now she starts a fresh conversation with a tight summary of the decisions made so far, instead of scrolling back and re-pasting everything. Output improves instantly, and a smaller desk costs less to run too (Lesson 9's metering, now a management habit).

That's optimisation, not brevity for its own sake. Include everything the task needs, and nothing it doesn't.

Where the picture breaks

A student still remembers pages after they slide off a real desk. The model doesn't: it has zero trace of off-desk material, ever. And a huge desk isn't used evenly. Material buried in the middle of a very long conversation can be under-weighted, even while it's technically still "on the desk". Bigger isn't automatically better. Well-managed is.

Now make the call

Kerry's hour-long care-plan chat has gone vague and repetitive. She has ten minutes before her next client. What does she do?

Apply it to your work

This week, at your desk

Next time a long AI conversation starts losing the plot, don't keep pushing through it. Write a three-line summary of what matters (decisions made, facts to keep, what you need now) and start fresh with it. Add this as a habit on your capstone checklist.

Open your AI checklist starter

Prove it to yourself

Two quick questions, untimed and retryable. They count toward ★ Mastered (completed + decision + self-check all correct). This is what stops "that made sense" evaporating by next week.

Explain it in your own words

Write two or three sentences for a colleague. Then compare your note against the rubric below. It checks ideas, not wording, and only you see what you write.

  • Names the context window as a fixed token budget
  • States that older material drops off; it doesn't fade
  • Connects this to a management habit: summarise or restart

Evidence & review — how we know what this lesson claims
Claim register for Lesson 10
ClaimTypeBasisReview risk
Every inference run is bounded by a fixed maximum token budget (the context window)How it worksArchitecture documentationLow
When the budget is exceeded, older content is dropped or compressed rather than the model "forgetting" graduallyHow it worksArchitecture documentationLow
Material positioned in the middle of very large contexts can be under-weighted ("lost in the middle")Research findingPrimary research literatureMedium
Specific context window sizes and per-product truncation/summarisation behaviourProduct behaviourProvider documentationHigh — reviewed quarterly
The desk simulatorTeaching deviceDeterministic toy demonstration, labelled illustrativeMedium
The open-book exam analogyUseful mental modelLimits stated in the lessonLow

Concept review due: January 2027. Product-behaviour claims: quarterly.

Where you are in the machine

You've stepped onto the scaffolding: the desk that everything else in this act gets built on. Select any layer to see its role, its lessons, and your live progress.

The machine, bottom to top: training data (2) → token pieces (3) → prediction engine (1, 4–9) → the desk (10–11) → senses (12) → live knowledge (13) → tools (14) → memory (15) → agent loops (16) → human judgement (17–21).

Next: what you put on the desk

The desk is set. What you put on it, and how you phrase it, changes everything. That craft is called prompting.