ficheLong chats, a quarter of the tokens

Draws old chat history onto page images and sends those instead of text. The model still reads all of it.

~74% fewer input tokens

On a 285,000-character Korean chat, 84k tokens per turn instead of 321k.

36 of 40 recall questions right, against 38 with the full text. No retrieval step that can miss a scene.

Pages are append-only, so the prompt prefix stays byte-identical and prompt caching keeps hitting.

1932 px PNG pages for Claude, vector PDF pages for Gemini. Standard APIs, no fine-tuning, no vector DB.