How it works

Every export becomes the same normalized messages. Then every action of the persona is replayed to the model as a tool call, so it learns how the person interacts, not just what they write, and answers the same way.

Actions as tool calls

What the persona did	Tool call in the history and in live answers
text bubble, maybe a reply or partial quote	send_message {text, reply_to?, quote?}
sticker	send_sticker {sticker_id, reply_to?}
GIF from the library, or a URL from an MCP tool	send_gif {gif_id | url, caption?}
photo, video, voice, audio, file	send_media {media_id, caption?}
forwarded post	forward_message {message_id}
reaction	react {message_id, emoji}
(recall older history)	search_history {query}

Several calls in one turn become several bubbles. The other side's messages become compact lines the ids refer to:

#52108 19:08 ↩ reply to #52107 (your: "grandmas lasagna 🍝"): You have made lasagna exactly zero times.
<sticker stk_156dbb381a>: a cat in a chef's hat, stirring a pot
reacted 😂 to #52107 ("grandmas lasagna 🍝")

The system prompt

Part
Who to be	estimate the persona's knowledge and style, and never know more than they would
Style profile	from the whole export: message length, capitalisation, punctuation, emoji and sticker use, reply and reaction habits, bubbles per turn, verbatim samples
Library	stickers, GIFs and media with how they were used
Notes	what you wrote about the person

The prompt is static for a chat: no clock, no counters. The model reads the time from the — day — separators and the HH:MM stamps, and each request is a byte-identical prefix of the next one, so provider prompt caches keep hitting.

Long chats

 history budget ───────────────────────────────────────────────▶
 ┌──────────────────────┬────────────────────────────────────────┐
 │ memory (≤ 350 words) │ newest turns, cut to 60% of the budget │
 └──────────────────────┴────────────────────────────────────────┘
   relationship, facts,    the cut point stays put for a long
   topics, running jokes   stretch, so the cached prefix does too

When the replayed history outgrows the budget, the next reply first folds the oldest turns (and the previous memory) into a short summary. Nothing is spent on import or when you open a chat. The style profile still comes from the whole export, and search_history reaches every old message. Token usage (prompt, cached, completion) is shown in { }.

Media library

The chat model reads text, so a sticker is just <sticker stk_…> until it is described. Describe with AI sends the images of the missing items to the media model, which must support vision.

• GIFs, video stickers and videos go as 4 frames in order, so the model sees the motion. This needs ffmpeg (brew install ffmpeg, apt install ffmpeg, winget install ffmpeg), otherwise the thumbnail is used.
• Redo describes one item again. Any description can be edited by hand.
• Descriptions belong to the export and serve all its chats.

See it yourself with the CLI: rooftop convert <export> --persona "Jonas Weber".