Performance Tips

Connection pooling

Use pool_min_size and pool_max_size to match your workload:


store = MemoryStore(
    "postgresql://...",
    pool_min_size=2,
    pool_max_size=20,
)

Increase ef_search for better recall at the cost of latency:


store = MemoryStore("postgresql://...", ef_search=200)

Disable with reranker_enabled=False to save ~10ms per recall if precision is acceptable:


store = MemoryStore("postgresql://...", reranker_enabled=False)

Install onnxruntime for 2-3x faster embeddings vs PyTorch:


pip install onnxruntime

Unforget auto-detects ONNX and uses it when available.

Run during low-traffic periods to avoid competing for database connections:


scheduler = ConsolidationScheduler(
    store,
    interval_seconds=3600 * 4,  # every 4 hours
    write_threshold=200,
)