Isolation is measured
MemQ returned leakage-free results in 100% of runs. Mem0 OSS and the keyword baseline each reached 67% in this corpus.
MemQ benchmark evidence
Published snapshotPublished retrieval results
In this 108-artifact snapshot, MemQ ranked the intended memory first in 100% of runs versus 58% for Mem0 OSS, with 13 ms average retrieval versus 2,511 ms.
Generated Apr 28, 2026 · 12 cases · 3 repetitions
Primary@1100%+42 points vs Mem0 OSS
Average retrieval13 ms193.2x lower than Mem0 in this harness
Leakage-free100%Across 3 isolated namespaces
Direct comparison
Retrieval quality and latency, side by side
Every percentage and timing below is read from the published snapshot. Longer quality bars are better; shorter latency bars are better.
Intended memory ranked first
Primary@1 across the same retrieval cases
MemQ MCP100%
Mem0 OSS58%
Keyword baseline58%
Average retrieval latency
Measured in the published harness; lower is better
MemQ MCP13 ms
Mem0 OSS2,511 ms
The in-process keyword baseline is excluded from the latency comparison because it does not make an equivalent service call.
Evidence ledger
Enough context to inspect the claim
The page keeps the benchmark envelope visible instead of turning a single number into an unsupported promise.
- Cases
- 12 7 retrieval families
- Namespaces
- 3 apollo, orion, zephyr
- Repetitions
- 3 Per provider and case
- Artifacts
- 108 Published result records
Answer quality is separate
The fixed-model answer corpus records 75% pass with MemQ context versus 0% with no memory across 12 cases.
The source stays downloadable
The generated JSON, summaries, case families, and repository remain linked so the headline can be checked against its source.
Method and limits
What these results do and do not prove
- The same 12 cases and 3 repetitions were used for each retrieval provider.
- The corpus covers disambiguation, freshness, incident, preference, procedural, project_memory, runbook.
- Primary@1, recall@K, MRR, namespace leakage, and service latency are calculated from the versioned run artifacts.
- Results describe this harness, dataset, configuration, and run date. They do not guarantee performance for every corpus, deployment, or workload.
Published sources
Read the evidence in its original format
Machine-readable results for automation, concise summaries for review, and the public harness for reproduction.