Question Answering on Crafter four-actor rollout robustness
59.3Average Exact Match (EM)S3Mem
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| S3Mem2026.04 | 59.3 | 148.64 | 54.4 | 61.82 | |
| Graph-NoReader2026.04 | 57.42 | 1,920.48 | 53.57 | 60.68 | |
| LightMem-adapted2026.04 | 56.37 | 111.94 | 53.85 | 61.98 | |
| A-MEM-inspired2026.04 | 52.7 | 142.53 | 46.43 | 58.4 | |
| MemoryOS-adapted2026.04 | 51.23 | 143.92 | 48.35 | 53.72 | |
| Vanilla RAG2026.04 | 46.43 | 788.46 | 42.58 | 49.57 | |
| No-Memory2026.04 | 32.25 | 462.73 | 31.05 | 32.78 |