Skip to main content

Storage Engines

Memorose combines several engines because no single storage model is enough for the full runtime.

RocksDB

Used for:

  • local durable key-value state
  • event and memory bookkeeping
  • task persistence
  • graph edge support structures
  • organization knowledge metadata

Lance

Used for:

  • embedding storage
  • vector similarity retrieval
  • efficient ANN-style search over memory units

Tantivy

Used for:

  • full-text retrieval
  • exact and keyword-heavy lookup
  • hybrid scoring alongside vector results

Why Hybrid Storage Exists

Memorose needs to support:

  • append-heavy ingest
  • structured task state
  • semantic retrieval
  • keyword retrieval
  • graph expansion
  • projected shared knowledge

That is why the project is a runtime with multiple cooperating storage paths rather than a single database wrapper.