Evidence Retrieval on Synthetic benchmark B=128 197 episodes, 907 queries
2.7Hit@2OSU-Mem
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| OSU-MemConstruction principle=overlap2026.06 | 2.7 | 13 | |
| Flat + MMR2026.06 | 2.2 | 8.5 | |
| Flat Vector2026.06 | 1.6 | 8.8 | |
| Coverage-Aware2026.06 | 1.4 | 8.6 | |
| Disjoint Hierarchy2026.06 | 1.4 | 8.4 | |
| Tree Abstraction2026.06 | 1.2 | 9.3 | |
| OSU-Mem (disjoint)Construction principle=disjoint2026.06 | 0.8 | 7.5 | |
| Budget-Trunc Replay2026.06 | 0 | 0.4 | |
| Sliding Window2026.06 | 0 | 0 | |
| Periodic Summary2026.06 | 0 | 4.1 |