A four-layer knowledge system that gives Claude persistent, queryable memory across sessions. Runs entirely on-host — Qdrant + llama.cpp + Python. No cloud, no API keys, no per-token cost on the embedding or memory layer.
This write-up walks through what it is, why it exists, the architecture, the phased build (with the bugs we hit and what they taught us), measured results, and how to set up your own.