A project by Virat Kaukuntla
Mnemo
Raise your own AI from scratch, on your own computer.
Mnemo is a self-learning AI built by Virat Kaukuntla. It starts with random weights and no vocabulary, builds its own tokenizer, reads a library and the live web, and trains while you leave it running. As it learns, it grows its own brain from 12 million to over 1 billion parameters without forgetting what it already knows. No pretrained weights are used: everything Mnemo knows, it learned on your machine.
What it does
- Its own decoder-only transformer in PyTorch (rotary positions, grouped-query attention, SwiGLU, RMSNorm) and its own byte-level tokenizer
- Grows in stages, 12M → 39M → 99M → 210M → 545M → 1.04B parameters, using a function-preserving expansion so nothing is lost
- Prompts up to 100,000 tokens and an unlimited, searchable memory of every page it reads
- A built-in web browser it can drive, plus Mnemo Code and Mnemo Studio
- Runs on macOS, Windows and Linux
Made by Virat Kaukuntla (Virat Reddy Kaukuntla), a high school software developer in Texas who also built VRK, Chroni, Animedle, Strata, AniRadar, Skelr, Pong Rivals, re://battlegrounds.
Last updated