Memory LanceDB

LanceDB-backed vector memory for high-volume embedding and retrieval workloads.

• LanceDB stores embeddings as Lance columnar format on disk — significantly faster for large collections (>100k vectors) than sqlite-vec.
• Use this backend when the default sqlite-vec backend shows degraded search performance under heavy indexing load.
• All standard memory operations (memory_search, memory_store, memory_get) work identically — the backend switch is transparent.
• Configure dbPath in the extension settings to point to the desired LanceDB directory (defaults to data/lancedb/).
• For auto-capture mode: the agent automatically stores conversation context chunks without explicit memory_store calls.
• For auto-recall mode: relevant memories are automatically injected into context at session start without explicit memory_search calls.
• When switching from sqlite-vec to LanceDB: run a memory backfill to migrate existing embeddings to the new backend.
• Monitor embedding batch health via /api/memory?action=embedding-status to confirm the LanceDB backend is active.