Update — ran a real bench on the live cluster (59 memories: 8 canonical facts × 3-4 paraphrases + 6 seeded contradictions + 20 distractors). Numbers:
duplicates per query (top-10): 0.9 → 0.0
top-result correct: 75% → 87.5%
11 consolidations in 80ms
conflicts detected: 0 of 6 seeded ← this one matters
Turns out conflict detection runs on graph edges, and /v1/remember doesn't auto-extract entities — so contradictions sit there invisibly until you explicitly call relate. That's a UX gap, not a missing feature, but it breaks the "drop memories in, get contradictions out" mental model. Filed as issues #1 and #2. Dataset + script + raw results: https://gist.github.com/spranab/49c618d3625dc131308227103af5.... Honest benches surface the kind of thing demos hide; thanks for pushing.
alt Hacker News
Update — ran a real bench on the live cluster (59 memories: 8 canonical facts × 3-4 paraphrases + 6 seeded contradictions + 20 distractors). Numbers:
duplicates per query (top-10): 0.9 → 0.0 top-result correct: 75% → 87.5% 11 consolidations in 80ms conflicts detected: 0 of 6 seeded ← this one matters Turns out conflict detection runs on graph edges, and /v1/remember doesn't auto-extract entities — so contradictions sit there invisibly until you explicitly call relate. That's a UX gap, not a missing feature, but it breaks the "drop memories in, get contradictions out" mental model. Filed as issues #1 and #2. Dataset + script + raw results: https://gist.github.com/spranab/49c618d3625dc131308227103af5.... Honest benches surface the kind of thing demos hide; thanks for pushing.