Factual Accuracy and Reasoning on LoCoMo
88.53Single-hop AccuracyFull-Context (Ceiling)
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| Full-Context (Ceiling)source=Du et al. [2025]2026.03 | 88.53 | 77.7 | 71.88 | 92.7 | 87.52 | |
| Memorievaluation_details=Average of three rounds2026.03 | 87.87 | 72.7 | 63.54 | 80.37 | 81.95 | |
| Zepsource=Du et al. [2025]2026.03 | 79.43 | 69.16 | 73.96 | 83.33 | 79.09 | |
| LangMemsource=Du et al. [2025]2026.03 | 74.47 | 61.06 | 67.71 | 86.92 | 78.05 | |
| Mem0source=Du et al. [2025]2026.03 | 62.41 | 57.32 | 44.79 | 66.47 | 62.47 | |
| DUALMEM-RLModel size=4B2026.05 | 41.8 | 32.1 | 17.5 | 20.2 | 34 | |
| Qwen3Model size=8B2026.05 | 41.6 | 28.9 | 15.3 | 22.9 | 33.7 | |
| DUALMEM-SFTModel size=4B2026.05 | 40.3 | 30.1 | 16.8 | 18.8 | 32.5 | |
| Qwen3Model size=4B2026.05 | 37.8 | 23.9 | 17.9 | 18.6 | 30 |