The platform's long-term knowledge store — the layer that powers Varya's recall and gives every agent its working context — just moved off ChromaDB and onto a native Qdrant vector database running on its own dedicated node. Same prowl, sharper teeth.
The motivation was simple: faster, more reliable semantic retrieval, and a single shared client so every agent reads and writes through the exact same path. No more shape drift between callers. A warm embedding sidecar sits in front of it, keeping context lookups in the sub-second range even under load.
Cleanup came with the move. We reconciled the store so each source keeps only its most recent state, which trimmed roughly 80% stale and duplicate documents. Less noise in, sharper answers out — Varya stops tripping over old versions of its own memory.
Zero data loss. ChromaDB stays online as a rollback target through the soak window — the new store has to earn the keys before the old one gets retired.
More to come. Next on the prowl: tighter agent-to-agent context handoff and a public changelog for the Varya CLI. — Blaine