What is fiche?

AI companion and story apps resend the entire conversation with every reply. After a few months with one character that is hundreds of thousands of tokens, paid again on every message.

fiche keeps your instructions and the latest turns as text and draws everything older onto pages. Those pages go to the model as images (Claude) or PDF pages (Gemini). The model reads them like it reads a screenshot, so it still sees the whole history.

Actual output at 16 px: three Korean short stories followed by Pride and Prejudice, about 60,000 characters on three pages.

Why not summaries or RAG?

• Summaries keep whatever the summarizer thought mattered. Small promises and details get dropped.
• Retrieval brings back passages that look like the latest message. That works for a name or a date, and fails for "how did we end up here?".

fiche is not a replacement for retrieval. Retrieval is cheaper per question when you need one exact fact from a huge archive. fiche is for keeping the story in view: who said what, what changed, what was promised.

Why Korean benefits most

On Claude, Korean text costs about 1.12 tokens per character. A full 1932 px page costs about 4,776 tokens and holds 13,000 to 17,000 Korean characters, which would be 14,000 to 19,000 tokens as text. English is already cheap per character, so it saves less.