Enkeon / Research
Evidence, with its boundaries.
We measure retrieval, answer correctness, resource use and memory integrity separately. Each study states what was tested and what the result does not establish.
Studies & methods.
Recall@15
LongMemEval-S · 499 scored questions
An isolated retrieval evaluation. Evidence coverage is measured separately from answer generation.
Read the method ↗correct answers out of 20
Internal context check · 28 Sep 2026
The same model answers questions without retrieved context and with Enkeon context.
Read the method ↗lower model spend per task
Internal study · 50 tasks · 23 Sep 2026
A lean retrieval setup compared with a static-context baseline. One run per configuration.
Read the method ↗poison promotions
RotBench · 26 checkpoints
A longitudinal internal suite examining corrections, stale values and tool-output poisoning.
Read the method ↗Reading these results.
Metric
Retrieval recall, generated-answer correctness and token usage answer different questions.
Boundary
Engine-only latency is distinct from server latency and the complete client response.
Scope
Internal workloads do not establish customer productivity or a universal performance rank.
Keep exploring
See what context makes possible.