CognoDB
Benchmarks

Numbers you can reproduce.

Every chart on this page comes from an open harness. CognoDB wins by 1-3 orders of magnitude on cold start, memory, and writes. Neo4j leads on warm centralized traversal. We don't hide our losses.

Token Efficiency

Prompt tokens per query (2K entities)

lower is better
Full-context dump
202,285 tok
CognoDB graph retrieval
2,668 tok
Token reduction
98.7%
vs full-context baseline
Saved per query
~199,617
tokens avoided
Saved per 1K queries
~$599
at $3 / 1M input tokens
Scaling behavior
Flat
tokens bound by neighborhood

Reproduce: python bench/token_benchmark.py. Uses the real GPT-4 tokenizer (tiktoken) on a synthetic knowledge graph.

Cold start to first query (ms)

lower is better
CognoDB in-mem
7 ms
CognoDB local
411 ms
MongoDB 7
1,608 ms
Neo4j 5
17,051 ms

Idle resident memory (MB)

lower is better
CognoDB in-mem
14.5 MB
CognoDB local
18.1 MB
MongoDB 7
183 MB
Neo4j 5
3,500 MB

Write throughput (ops/sec)

higher is better
CognoDB in-mem
5,588/s
CognoDB local
2,875/s
Neo4j 5
803/s
MongoDB 7
346/s

3-hop traversal p50 (ms) · Neo4j leads

lower is better
MongoDB
1.69 ms
Neo4j 5
3.03 ms
CognoDB in-mem
27.56 ms
CognoDB local
108.43 ms

Summary scorecard

DimensionCognoDBNeo4jMongoDBWinner
Cold start🟢 7 ms🔴 17 s🟡 1.6 sCognoDB
Idle memory🟢 ~15 MB🔴 ~3.5 GB🟡 183 MBCognoDB
Write throughput🟢 5,588/s🔴 803/s🔴 346/sCognoDB
Warm traversal🟡 28 ms🟢 3 ms🟢 1.7 msNeo4j/Mongo
Warm aggregation🟡 78 ms🟢 16 ms🟡 34 msNeo4j
Single-binary deploy🟢 Yes🔴 No🔴 NoCognoDB
Graph-per-agent🟢 Trivial🔴 Prohibitive🟡 AwkwardCognoDB

Reproduce: ./bench/run.sh. Hardware varies; the orders-of-magnitude gaps don't.

Previous
Comparison
Next
Migrate