I build memory and agent systems for LLMs, mostly around long-running role-play: systems that stay persistent, inspectable, and reproducible beyond a single context window.
I like architectures where the model can be probabilistic without making the entire application unpredictable.
- NMOS: local, auditable long-term memory for PocketRisu / RisuAI chats. Immutable history, lexical and optional semantic recall, bounded injection with provenance.
- MARP: multi-agent pipeline for RisuAI that analyzes world, plot, and characters in parallel before the main model replies.
Python · TypeScript · Rust
Most of what I build starts with "I wonder if this would work..."
and gets slightly out of hand from there.