AI Symbiont

Frontier models know everything except your codebase. ai-symbiont teaches it. You corrected your coding agent last week. Today it makes the same mistake again. THE RESULT In sqlglot, a real open-source project, one rule lives only in the maintainers' code reviews. Muse Spark broke it in 7 of 10 runs without symbiont. With symbiont it broke it in 0 of 10 (p = 0.003). Every run passed the tests. This is one repo, one rule, one task, and it was pre-registered. WHAT IT IS ai-symbiont is a plugin for Meta's Muse Code. It reads your past agent sessions, your corrections and your team's pull-request reviews. It turns them into lessons. Just before the agent repeats a mistake, it shows one plain sentence, often in the reviewer's own words. Nothing leaves your computer except the agent's normal calls to Meta's API. HOW THE META MODELS FIT - Muse Code is the host. Hooks bring notes into the agent's loop. One command, /symbiont onboard, learns from your history and trains the checker. - Muse Spark is the agent. It also reviews candidate lessons and keeps the ones that hold up. Spark itself is not changed. symbiont makes no model calls of its own. - Llama 3.2 1B (open weights) is the checker. A LoRA is trained on each repository's own history, on your machine. It must agree before any note is shown, because a wrong note is worse than none. On 89 open-source projects it had never seen, it showed 0 wrong notes in 501. Plain similarity search showed 9 wrong. - Muse Glimmer 30B (open weights) is an early test and my next step. With symbiont's lessons in its prompt it solved 6 tasks against 2 without them. In a separate test on Llama 3.2 1B, lessons trained into the weights were recalled 88% of the time. Getting Glimmer to use them while it codes is still open. MORE EVIDENCE All of it is pre-registered, and every limit is published. - Corrected once, never repeated: 16 rule violations without symbiont, 0 with it, across 8 new tasks. - A home robot in Meta's Habitat learned to set the table from its own two failures. It went from 0 of 5 to 4 of 5. Giving it every lesson up front also reached 4 of 5, so the win is the lessons it learned. It also stopped breaking house rules it was told once, from 30 to 0. - Muse Spark plays Freedoom via ViZDoom. It proposed six rules for itself. An automatic test kept one. That rule raised its score across 240 new games. TRY IT Install the plugin and run /symbiont onboard in Muse Code. Or click through the keyless demo on the project page. Or run a live Muse Code session in Docker with your Meta API key. Built from scratch during the hackathon, Oct 26 to Nov 4, 2026. Github Page Showcase, recordings and paper: https://smalls257.github.io/ai-symbiont/ Code: https://github.com/smalls257/ai-symbiont