Skip to main content
TL;DR: Memvid reduces retrieval errors by 50-66% compared to leading vector databases while delivering 4x better accuracy-per-millisecond than alternatives.
We benchmarked Memvid against four popular vector databases on the Wikipedia dataset (39,324 documents, 2,500 queries). The results speak for themselves.

Headline Results

92.7%

Top-1 AccuracyMemvid finds the correct document on the first try 92.7% of the time, 8.5 to 14.5 points higher than any competitor.

50-66%

Error ReductionMemvid reduces retrieval errors by 50-66% compared to Chroma, Weaviate, LanceDB, and Qdrant.

4.2x

Accuracy per msMemvid delivers 4.2x more accuracy per millisecond than Chroma, the best quality-to-latency ratio.

0.95

MRR ScoreMean Reciprocal Rank of 0.95. The correct result consistently appears in the top 1-2 hits.

Accuracy Comparison

Top-1 Accuracy (Higher is Better)

The most important metric: does the system return the correct document first?
SystemAccuracy@1vs Memvid
Memvid92.72%baseline
LanceDB84.24%-8.5 pts
Qdrant84.24%-8.5 pts
Weaviate80.68%-12.0 pts
Chroma78.24%-14.5 pts

Error Reduction

Another way to look at accuracy: how often does the system get it wrong?
SystemError RateMemvid Reduces Errors By
Chroma21.76%66% fewer errors
Weaviate19.32%62% fewer errors
LanceDB15.76%54% fewer errors
Qdrant15.76%54% fewer errors
Memvid7.28%baseline
Memvid returns the wrong answer 3x less often than Chroma.

Accuracy at Different k Values

System@1@3@5@10
Memvid92.72%96.96%97.56%98.12%
LanceDB84.24%92.72%94.60%96.24%
Qdrant84.24%92.72%94.60%96.24%
Weaviate80.68%88.52%90.16%91.52%
Chroma78.24%85.80%87.28%88.68%

Ranking Quality

MRR (Mean Reciprocal Rank)

MRR measures how high the correct result appears in the ranking. A score of 1.0 means perfect top-1 placement every time.
SystemMRRInterpretation
Memvid0.949Correct result typically #1
LanceDB0.888Correct result typically #1-2
Qdrant0.888Correct result typically #1-2
Weaviate0.849Correct result typically #2
Chroma0.823Correct result typically #2

NDCG@10 (Normalized Discounted Cumulative Gain)

NDCG measures overall ranking quality across the top 10 results.
SystemNDCG@10
Memvid0.967
LanceDB0.929
Qdrant0.929
Weaviate0.886
Chroma0.858
Only Memvid achieves >0.94 MRR and >0.96 NDCG@10 in this benchmark. Memvid consistently surfaces the correct result in the top 1-2 hits with near-perfect ranking quality.

Latency Performance

Query Latency (Lower is Better)

Systemp50p95p99QPS
Weaviate5.3ms7.1ms7.9ms180
Memvid16.0ms17.4ms19.7ms61
LanceDB16.0ms17.5ms19.3ms61
Qdrant28.0ms30.4ms31.4ms36
Chroma55.6ms61.0ms65.2ms18

Cold Start Time

SystemCold Start
Memvid0.5ms
Chroma66.3ms
Qdrant71.8ms
LanceDB72.4ms
Weaviate147.7ms
Memvid cold-starts 130-300x faster than alternatives. This matters for serverless deployments and edge computing where startup time is critical.

The Efficiency Frontier

Accuracy per Millisecond

We compute accuracy divided by p95 latency to measure quality-per-latency:
SystemAccuracy@1p95 LatencyAccuracy/ms
Memvid92.72%17.4ms0.053
Weaviate80.68%7.1ms0.114
LanceDB84.24%17.5ms0.048
Qdrant84.24%30.4ms0.028
Chroma78.24%61.0ms0.013
Memvid delivers 4.2x more accuracy per millisecond than Chroma and leads all systems except Weaviate (which sacrifices 12 points of accuracy for speed).If accuracy matters, Memvid is the clear winner. If you can tolerate 12% worse results, Weaviate is faster.

The Frontier Chart

Accuracy@1 (%)

 93% │                    ★ Memvid

 85% │        ● LanceDB   ● Qdrant

 81% │    ● Weaviate

 78% │                              ● Chroma

     └────────────────────────────────────────
         5ms   17ms  28ms           61ms
                     p95 Latency →
Memvid sits alone at the accuracy frontier. No other system achieves >90% accuracy at any latency.

Storage Efficiency

Single-File Advantage

Memvid is unique among memory systems because everything lives inside a single portable file; (data, embeddings, indices, and metadata) There are no sidecar files, external indexes, database directories, or hidden dependencies. One file contains the entire memory: easy to move, copy, version, share, or embed into an agent.
SystemStorageCompressionBytes/Doc
LanceDB213 MB0.71x5,428
Qdrant212 MB0.72x5,396
Memvid508 MB0.30x12,911
Weaviate1,009 MB0.15x25,646
Chroma1,025 MB0.15x26,068
Memvid’s higher bytes-per-document value reflects its richer internal structure: embedded indices, a write-ahead log, a time index, and the metadata required for hybrid semantic + keyword + time-travel search. Instead of scattering these components across multiple files or services, Memvid packages the entire memory system into a single .mv2 file, delivering portability and simplicity that traditional systems can’t match.

Methodology

Dataset

  • Wikipedia: 39,324 documents
  • Queries: 2,500 natural language queries with known correct answers

Systems Tested

  • Memvid v2 (hybrid search mode)
  • Chroma 0.4.x (default HNSW)
  • LanceDB (default IVF-PQ)
  • Qdrant (default HNSW)
  • Weaviate (default HNSW)

Metrics

  • Accuracy@k: Fraction of queries where the correct document appears in top-k results
  • MRR: Mean Reciprocal Rank (1/position of first correct result)
  • NDCG@10: Normalized Discounted Cumulative Gain at 10
  • Latency: Query time in milliseconds (p50, p95, p99)
  • QPS: Queries per second throughput

Environment

  • Apple M-series Mac
  • All systems using default configurations
  • Same embedding model across all systems
  • Each query executed 1x (no caching)

Key Takeaways

Best Accuracy

Memvid achieves 92.7% top-1 accuracy, 8.5 to 14.5 points higher than any competitor.

Lowest Error Rate

Memvid reduces errors by 50-66% compared to leading vector databases.

Best Ranking

Highest MRR (0.95) and NDCG@10 (0.97). The correct result consistently appears first.

Fastest Cold Start

0.5ms cold start, 130-300x faster than alternatives for serverless deployments.

Run the Benchmarks Yourself

The benchmark suite is open source. Run it on your own hardware:
git clone https://github.com/memvid/memvid
cd memvid/benchmarks/python
pip install -r requirements.txt
python run_benchmarks.py
Results are saved to results/results.json.

Next Steps