A clone is not “an assistant that describes a person” — it is a digital cast of their speech: reply length, favourite words, emoji, turns of phrase, the habit of answering in a burst of short messages. Upload a chat export, pick a participant, and the agent starts talking in their voice.
The export is parsed RIGHT IN YOUR BROWSER: the zip is unpacked locally, HTML and result.json are read on your machine. A script deterministically computes a compact style profile — reply statistics, word and emoji frequencies, frequent phrases, a small verbatim sample. Only that profile (tens of kilobytes) is sent to the server, never the chat history.
This is not only about privacy: frequencies are COUNTED rather than invented by a model, and one brain call over the profile replaces dozens of calls over chunks of history.
The profile is stored with the agent: a clone can be REBUILT without re-uploading files and TOPPED UP from new chats — new data is merged with the old deterministically.
Voice fidelity rests on three layers: a locked internal rulebook for clones, clone-aware memory and knowledge (the double does not read out a biography), and “meaning from usage” — catchwords land where the person actually used them.
Nowhere. The export files are parsed in your browser, and only a compact style profile is sent to the server: numeric statistics, word and emoji frequencies and a small sample of phrases. The conversation itself is never uploaded to Corpis servers.
In Telegram Desktop open the chat, use the menu in the top right corner and choose “Export chat history”. Pick JSON or HTML; media is not needed — the clone only uses text. Drop the resulting folder or .zip onto the clone page.
Like a person — that was the main fix of the pipeline. The clone keeps the original reply length and rhythm, uses their catchwords in the right places, does not slip into “how else can I help you” and does not recite a biography unless asked.
Technically yes, a shared chat export contains both sides. The legal and ethical side of impersonation stays with the agent owner: by publishing a clone you take responsibility for the consent of the person whose voice you reproduce.