Skip to main content

Try a simple example

What can AI remember?

Add one message to a tiny conversation. Watch which detail no longer fits.

Prepared teaching exampleNo typing needed

Add one more message.

The blue gate is in the selected context.

Included in this request

12 of 12 illustrative blocks
Instructions2 blocks
Question2 blocks
Room for an answer2 blocks
The question we want to answer

Which gate did we choose?

  1. Meeting place2 blocks

    Let’s meet at the blue gate.

  2. Weather2 blocks

    It might rain. Bring a coat.

  3. A snack2 blocks

    I’ll bring apples for the walk.

Outside this request

Every history card fits so far.

Illustrative blocks, not measured tokens. This demo drops the oldest whole cards; real apps use different policies.

The worked example

A 12-block window reserves two blocks for instructions, two for the question and two for an answer. Six remain for three 2-block history cards: the blue gate, the weather and a snack. Add a fourth 2-block card about a book. Under this demo’s oldest-first rule, the blue-gate card moves outside the request. Restating the gate adds it back as a new card. Real apps need not follow this rule.

Still in the chat does not mean included now.

In this example, the fourth card pushes the gate detail out of the request. The conversation still shows it, but this selected context does not contain it.

Bring important details back into the request

Restate the facts and constraints that matter now. A long conversation alone is not a reliable checklist for the next answer.

Go deeper

What is a context window?

It is the bounded information a model can use for one response. Input and generated output share limits defined by the model and system. Our reserved blocks make those competing uses visible; they are not measured token counts.

Do real apps always drop the oldest message?

No. That is this demo’s explicit policy. Apps may select messages, summarize, retrieve older material, stop at a limit, or reject an oversized request. A saved chat transcript and a model’s current input are different things.

Will the model reliably use everything that fits?

No. The Lost in the Middle study found that tested models could use long-context information unevenly depending on where it appeared. Space to include a detail is different from reliable use of that detail.

Explore the original sources

Primary references. These explain the mechanisms; the examples here are authored teaching simulations.

Try a related question

Follow your curiosity. Each experience stands on its own.

How does AI read words? →Look at the pieces a model works with.How does AI find evidence? →Choose a source that actually supports the answer.

What would you like to try next?

Choose another small example, or return to this one whenever you need it.

See all the examples