The lab
Interactive · drag the conversationInteractive · tap through the conversation

The illusion of memory

Watch a long AI chat lose an early fact — then see why asking again cannot bring that fact back.

The problem: a chat can seem to remember everything, then suddenly lose a name, instruction or decision you gave it earlier. In real work, that missing detail can change an answer while the conversation still looks intact.

Narrated by a synthetic voice — not a recording · ~0.7 MB

Scroll to the experiment, then drag ↓Scroll normally, then tap through the facts ↓
The setup

An AI does not hold a whole chat in a human-style memory. For each reply, the chat system places a bundle of recent words in front of it to read again.

That bundle has a size limit. Think of it as a fixed reading space: when new messages fill it, older words can fall outside. Engineers call this limited reading space the context window. Some chat products also fetch or summarise older material, but the AI can use only the text the product puts back in front of it for that reply.

WHAT IS ILLUSTRATIVE — The window, four story facts and reply box below are a teaching toy. The reply is a scripted check of four facts, not a live AI. Real limits count small text pieces (engineers call them tokens), and products handle old material in different ways. The real point shown here is narrower: text outside the material supplied for a reply cannot shape that reply.

Below, a story is mid-conversation and the AI's reading space is already full.

Tap the conversation to move one fact past the line. Tap through all four facts; scroll the page normally between sections.


01 · Watch a fact leave

What happens when a conversation grows too long?

On the right is the toy reading space, filled with facts about a story.

Drag the conversation upward.Tap the conversation to push the next fact out. Keep an eye on the reply at the bottom as each fact crosses the line.

Your dragEach tap stands in for the chat growing: every new message pushes the oldest words toward the edge.

This is not the AI searching a permanent memory and coming up blank. Once a fact crosses the tangerine line, its words are no longer in the reading space used for the reply.

All four facts are outside the window now. The AI isn't refusing — the words simply aren't in front of it.

02 · Asking questions

Can it answer about what just fell out?

Try it — ask about Elara's name, her city, her companion, or what she carries.

Your question changed which fact the toy checked — not which facts were still available.

03 · A shorter recap

Can a summary save the past?

One common repair is to shorten earlier messages into a recap and keep it at the top. Engineers call this compression. A shorter recap saves space, but it must choose which details to keep.

Most of her came back — name, city, droid. The silver wrench didn't. The shorter recap chose what to keep before you knew which detail you would need.

Reading-space edge Full

Drag upTap to advance
Pinned recap: Elara is an engineer from New Phoenix. She is travelling with a droid named P-3.
Let's write a story. Our protagonist's name is Elara.
Got it. Elara is our protagonist.
She's an engineer from the submerged city of New Phoenix.
A submerged city engineer. Sounds atmospheric!
She never goes anywhere without her father's silver wrench.
The silver wrench is added to her inventory.
Her only companion is a rusted surveyor droid named P-3.
P-3 the surveyor droid is on board.
They set out at dawn.
The water was calm as they departed...
They encountered a storm.
Dark clouds gathered quickly...
P-3 got struck by lightning.
Sparks flew from its chassis...
They found a cave to hide in.
The cave was dark and damp...
What to carry into work

A long chat is not proof that every earlier detail is still available to the AI.

What to do differently: when a long chat loses a name, rule or decision, put the exact fact back into your latest message and ask the AI to confirm it before relying on the next answer. For important work, keep the source outside the chat too; a recap can silently drop the detail you later need.

Narration transcript (synthetic voice)

0 · Intro (plays when you turn sound on)

You're in the lab. A long AI chat is about to lose an early fact. You'll push that fact beyond the text the AI can read, then see what brings it back — and what still gets lost.

The Setup

(The voice stays silent here — the notes on this part of the page are yours to read.)

1 · Watch a fact leave

Choose the conversation and move it past the tangerine line, one fact at a time.

1b · After a fact crosses the line

A fact crossed the line. Its words aren't in the AI's reading space now, so the reply below just lost them.

2 · Asking questions

Now ask it something yourself. Try her name, her city, or what she carries.

3 · A shorter recap

Try the shorter recap. Tap the pin button, then read the reply again.

3b · After you pin the recap

Most of her came back. The wrench didn't. The recap chose what to keep before you knew which detail you'd need.

4 · What to carry into work

One habit to take with you. When a long chat loses a detail, put the exact fact back into your latest message before you rely on the answer.

RK · a quiet place for long reading