Mnemosyne
A local-first memory OS for macOS that gives AI agents persistent, encrypted, versioned memory — running the models on your own Apple Silicon.
Overview
Agent memory shouldn't live on someone else's server. Mnemosyne is an encrypted, on-device knowledge graph that ingests your notes, mail, calendar and meetings, indexes them locally, and lets you query or act on them by voice — with every model running in-process on Apple Silicon.
I built it in Swift 6 with strict concurrency: about 145,000 lines in three months, solo. The part I'm proudest of is a provenance type — every value carries its origin on a five-level trust lattice, and reading it takes a named, greppable unwrap, so prompt-injection and declassification become things the type system can audit. Dangerous actions are gated below the authoring layer at a single chokepoint no chain author or injected prompt can opt out of.
Highlights
- On-device LLM inference and LoRA fine-tuning via MLX; Metal-accelerated vector search.
- Encrypted append-only content log plus a SQLite manifest — versioned and auditable.
- An honest "shipped vs. aspirational" ledger in the repo. Breadth without that is theater.