audrey-memory-benchmark-artifacts / memorygym-current-summary.json
Evilander's picture
Publish before-after MemoryGym benchmark rerun
f303647 verified
{
"generatedAt": "2026-05-01T04:41:01.251Z",
"beforeRun": "mg-2026-05-01T04-19-26-779Z-17400-c54ca52e",
"afterRun": "mg-2026-05-01T04-38-49-773Z-3032-2a8ec0b6",
"beforeGeneratedAt": "2026-05-01T04:19:26.858Z",
"afterGeneratedAt": "2026-05-01T04:38:49.849Z",
"suite": "MemoryGym Core",
"patch": "Audrey recall now requires all requested tags when multiple tag filters are provided.",
"audreyBefore": {
"scorePercent": 34.3,
"hitRatePercent": 42.9,
"p95RecallLatencyMs": 3.739
},
"audreyAfter": {
"scorePercent": 57.5,
"hitRatePercent": 85.7,
"p95RecallLatencyMs": 3.794
},
"delta": {
"scorePoints": 23.2,
"hitRatePoints": 42.9,
"p95RecallLatencyMs": 0.055
},
"adapters": [
{
"name": "typed-semantic",
"scorePercent": 72.7,
"hitRatePercent": 100,
"p95RecallLatencyMs": 0.368,
"precision": 0.34523809523809523,
"recall": 1,
"mrr": 0.7857142857142857,
"ndcg": 0.8507551301085979,
"contaminationPenalty": 0.4523809523809524
},
{
"name": "hybrid",
"scorePercent": 72.7,
"hitRatePercent": 100,
"p95RecallLatencyMs": 0.25,
"precision": 0.34523809523809523,
"recall": 1,
"mrr": 0.7857142857142857,
"ndcg": 0.8507551301085979,
"contaminationPenalty": 0.4523809523809524
},
{
"name": "audrey-http",
"scorePercent": 57.5,
"hitRatePercent": 85.7,
"p95RecallLatencyMs": 3.794,
"precision": 0.29761904761904756,
"recall": 0.8571428571428571,
"mrr": 0.4642857142857143,
"ndcg": 0.567769535513557,
"contaminationPenalty": 0.49999999999999994
}
],
"remainingGaps": [
"Current-belief ranking still fails: old workspace memory outranks the current workspace update.",
"Context routing now finds the expected Audrey/RepoPulse memories, but often ranks a project distractor first.",
"Contamination penalty remains high; next work should harden context-aware ranking and current-belief precedence."
],
"claimBoundary": "This is a current MemoryGym local benchmark artifact after an Audrey patch, not an official AMB, LoCoMo, LongMemEval, Papers with Code, or CodeSOTA leaderboard score."
}