This was an early experiment in GPU-accelerated codebase memory via embeddings. It's been superseded by graymatter — lighter, faster, and CPU-only.
| Approach | Mnemotech: GPU embeddings (LanceDB) | graymatter: FTS5 + LYT (SQLite) |
|---|---|---|
| Hardware | Requires GPU for indexing | CPU-only, runs anywhere |
| Token Savings | ~50-70% | 70-98% |
| Search Speed | ~100-500ms | <10ms |
| Complexity | High (embeddings, vectors, GPU) | Low (SQLite, text search) |
| Maintenance | ❌ No longer maintained | ✅ Actively developed |
Migration is easy: graymatter uses the same Obsidian vault structure.
Just install graymatter and run gm.py sync — your existing notes work immediately.
graymatter is the evolution of everything Mnemotech tried to do — without the GPU requirement.