Vector Search

Dense embedding-based similarity search using sqlite-vec for efficient nearest-neighbor queries on episodic memory.

How It Works

1. Embed: Query text is converted to a 384-dimension vector via BAAI/bge-small-en-v1.5 (loaded through fastembed)
2. Search: sqlite-vec performs nearest neighbor search on int8 quantized vectors stored in the vec_episodes table
3. Score: Cosine similarity between query and stored vectors
4. Rank: Results ordered by similarity (highest first)

Embedding Model

Mnemosyne uses BAAI/bge-small-en-v1.5 exclusively, loaded via the fastembed library:

The model runs entirely locally — no OpenAI API key or network calls are required for embedding generation.

Vector Storage

Vectors are stored in a sqlite-vec virtual table:

-- Internal sqlite-vec table (created automatically)
CREATE VIRTUAL TABLE vec_episodes USING vec0(
  embedding int8[384]
);

Usage via recall()

Vector search is performed automatically as part of the recall() method. There is no separate vector_search() API:

from mnemosyne import Memory

mem = Memory()

# recall() internally performs vector + FTS5 hybrid search
results = mem.recall("deployment strategy", top_k=5)

for r in results:
  print(f"Score: {r['score']:.3f} — {r['content'][:80]}")

Performance Characteristics