Long-context memory-based conversation reasoning on BEAM 100K (test)
59Pass@0.5 RateSelfMem
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| SelfMemScale=100K2026.07 | 59 | 50.4 | 0.984 | 36.2 | 1,932 | 0 | |
| RAGScale=100K2026.07 | 42 | 33.9 | 0.551 | 1.7 | 1,548 | 5,656 | |
| ReadAgentScale=100K2026.07 | 37 | 31.4 | 2.94 | 1 | 1,600 | 490 | |
| LoCoMoScale=100K2026.07 | 36.5 | 30.6 | 2.904 | 1.1 | 1,587 | 580 | |
| MemGPTScale=100K2026.07 | 36.2 | 29.1 | 6.434 | 1 | 1,582 | 450 | |
| Mem0Scale=100K2026.07 | 35.8 | 28.2 | 1.697 | 34.1 | 2,160 | 2,912 | |
| A-MemScale=100K2026.07 | 35.2 | 30.2 | 3.116 | 0.7 | 1,606 | 580 | |
| MemoryBankScale=100K2026.07 | 34.5 | 28.4 | 0.609 | 2.1 | 1,602 | 6,212 | |
| Full contextScale=100K2026.07 | 25.8 | 20.5 | 2.427 | 87.7 | 1,559 | 0 | |
| CompressionScale=100K2026.07 | 22.5 | 17 | 2.83 | 85 | 1,726 | 0 |