wtf is fractale?

Fractale is a usage kit for a type of language model that remembers what it reads in an unusual way. Instead of keeping the entire document in a growing prompt like a typical chatbot does, a Fractale model reads a long document one page at a time and is only allowed to keep 8 sticky notes about what it has seen. After each page it writes one note in its own internal shorthand, and once all 8 notes are full, the oldest one gets removed to make room. The model never goes back and rereads earlier pages. Instead, those notes plug directly into the model's own internal workings, so the model thinks through its memory rather than reading it back as text. This repository provides the tools to actually use a Fractale model: loading it, feeding it text, generating continuations, and treating its memory as something you can save to a file, restore later, reset, or swap between different documents. Because the model carries this compact memory state between calls instead of a growing prompt, using it works differently from a normal large language model, and this kit handles that loop for you. The included model is a 386 million parameter base model that has only been pretrained, so it does not have chat abilities and should not be expected to quote text word for word. Its memory captures the general gist, tone, and structure of what it read rather than an exact copy. The repository includes runnable demo scripts that compare what the model predicts with its memory turned on versus turned off, plus a demo that swaps memory banks between two different documents to see how predictions change. The project requires Python and is installed by cloning the repository and running pip install. It is released under the MIT license, and the underlying training code and research live in a separate companion repository.