ConvoMem
2,847 questions · Six evidence categories
99.6% on user evidence
CTX RESEARCH / BENCHMARKS
Longer conversations.
Compact context.
Useful memory should bring back what matters. Explore CTX’s answer quality, context efficiency, and retrieval speed.
ConvoMem
92.3%
Accuracy2,847 questions
Six evidence categories
Six evidence categories
LoCoMo
88.6%
Answer accuracy1,540 questions
4 categories · Excluding adversarial
4 categories · Excluding adversarial
01 / CONTEXT EFFICIENCY
Useful memory.
Compact context.
CTX keeps the context it brings back compact, averaging under 5K context tokens across both benchmark runs.
Under 5K average context tokens.
API-reported benchmark averages.
ConvoMem4,322 average context tokens
LoCoMo4,825 average context tokens
These are historical CTX-run evaluations, not independent certification or a guarantee for every workload. Scores depend on dataset, prompts, models, and scoring scope. Sequential retrieval latency excludes answer generation, client/network latency, cold starts, and concurrent-load behavior; it is not a service SLO.
02 / ANSWER QUALITY
The right details.
Across different demands.
Facts, preferences, changing information, and connections across conversations. See how CTX performs in each benchmark category.
Answer accuracy
LoCoMo
1,540 questions · 4 categories · Excluding adversarial
93.7% on single-hop questions
03 / RETRIEVAL SPEED
Context, without the wait.
Fast recall, backed by CTX’s production database.
Sequential requests · Server-side retrieval
Excludes answer generation
55 msMedian / p50
65 msp90
04 / EVALUATION NOTES
Behind the results.
CTX-run evaluations, with each benchmark’s scoring method and scope.
ConvoMem
Correct-answer accuracy across 2,847 questions, including abstention.
LoCoMo
1,365 of 1,540 answers judged correct across single-hop, multi-hop, temporal, and open-domain questions. Adversarial questions are excluded from this displayed score.