I’ve been working on Mati, a personal AI that runs entirely locally with encrypted memory. My goal was to model memory as a continuous process (a phase field) rather than a static database. I wanted to share the architecture and the honest limitations. For example, in my tests, field-based retrieval scored 0.036, while lexical retrieval scored 0.838. I document all of this, including the negative results, in the ARCHITECTURE.md. It’s built for CPUs, respects privacy by design, and is Apache-2.0. I’d love feedback from this community.(https://github.com/vml520/Mati)
Top comments (0)