elixirtrials/rune

Rune is a research project exploring whether LoRA weight space can serve as persistent episodic memory for coding agents, encoding procedural knowledge directly into model parameters rather than relying on context windows. It proposes a Doc-to-LoRA hypernetwork that converts coding trajectories into composable adapters, combined with recursive distillation and a hierarchical adapter library that grows with each solved task, enabling local Small Language Models to accumulate coding competence over time without cloud dependencies.