Benchmarks
Real benchmarks comparing Memvid to leading vector DBs, tested on 40k Wikipedia articles and 2,500 queries.
Last updated: Nov 15, 2025 · New benchmarks coming soon
State-of-the-Art Performance in Retrieval
Accuracy@1 improvement on Wikipedia benchmark with 40k articles and 2,500 queries.
Memvid Highlights
Best-in-class retrieval accuracy on Wikipedia benchmark
Best in classMedian query latency across 2,500 queries
Sub-20ms responsesTime to first query - instant availability
Instant startupMean Reciprocal Rank - ranking quality metric
Best overall rankingThroughput on 39K document corpus
High throughputNormalized Discounted Cumulative Gain
Highest search qualityFull Comparison
| Metric | Memvid | Chroma | LanceDB | Qdrant | Weaviate |
|---|---|---|---|---|---|
| Accuracy@1 | 92.7%✓ | 78.2% | 84.2% | 84.2% | 80.7% |
| Accuracy@10 | 98.1%✓ | 88.7% | 96.2% | 96.2% | 91.5% |
| Latency P50 | 16.0ms | 55.6ms | 16.0ms | 28.0ms | 5.3ms✓ |
| Latency P99 | 19.7ms | 65.2ms | 19.3ms | 31.4ms | 7.9ms✓ |
| Cold Start | 0.5ms✓ | 66.3ms | 72.4ms | 71.8ms | 147.7ms |
| MRR | 0.949✓ | 0.823 | 0.888 | 0.888 | 0.849 |
| QPS | 61.1 | 17.8 | 61.3 | 35.7 | 180.4✓ |
Storage Efficiency
Storage size for 40k Wikipedia articles with embeddings
Memvid stores everything in a single portable file — no database server required
Test Environment
Run your own benchmarks: cd benchmarks/python && python run_benchmark.py
Across every major retrieval metric, Memvid consistently outperforms leading vector databases, delivering higher accuracy, lower latency, instant cold starts, and dramatically better overall ranking quality. And it does all of this without a server, without pipelines, and inside a single portable file. These benchmarks reflect what developers see in real-world use: faster answers, better recall, and a simpler memory stack that just works.