AgentDB Performance Optimization

Enable binary quantization by setting quantizationType to 'binary' in createAgentDBAdapter, which reduces memory 32x (e.g., 1M vectors from 3GB to 96MB) with roughly 2-5% accuracy loss. Scalar quantization offers a 4x reduction with higher accuracy.

AgentDB automatically builds HNSW indices, delivering searches around 100µs versus 15ms for linear scans. You can tune hnswM, hnswEfConstruction, and hnswEfSearch parameters, and add an in-memory cache to bring repeated retrievals under 1ms.

Binary quantization gives 32x memory reduction with 95-98% accuracy, scalar gives 4x reduction with 98-99% accuracy, and product quantization gives 8-16x reduction with 93-97% accuracy. Choose based on dataset size and accuracy requirements.

Yes. Combine binary quantization with a small cache (around 100 entries) and low HNSW connections (M=8) to run roughly 100K vectors in about 10MB of memory while keeping search under 100µs.

Aggressive quantization or low HNSW search quality reduces recall. Switch from binary to scalar quantization, or raise hnswEfSearch to 200, to recover accuracy at the cost of some speed and memory.

Use batch inserts instead of individual insertPattern calls. Batching 100 vectors takes about 2ms versus roughly 1 second for sequential inserts, a 500x improvement according to the documented benchmarks.