Tokens
Text is converted into tokens
The Context Window
The context window is a fixed-size box
But your subscription tier decides how much you actually get
Multi-turn Conversation
Every message takes up space in the box
What Happens When It's Full?
Older messages
silently removed
Middle of conversation
summarised
✋
Session limit reached.
Start a new chat.
Chat ends entirely
Providers handle this invisibly — and differently
Hidden Instructions
Root
Cannot be overridden. Prevents catastrophic harm.
System
Set by provider. Varies by surface & user type.
Developer
API-level instructions from app builders.
These layers are already inside the box before you type
Long-Term Memory
Past chats and stored user data pre-fill the box
Document Uploads
Under 110k tokens — full document included
Over 110k — the remainder is only searchable, not seen
Reasoning Tokens
Reasoning tokens: temporary, invisible, discarded
The model is never operating on
what you think it is.
Should we demand more transparency?