Long-term Dialogue Memory Management on LoCoMo
42.77F1MGRetrieval
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| MGRetrievalBackbone=Qwen2.5-14B and Qwen3-14B, Evaluation Protocol=Weighted average, pyramid layers=4, max iterations=42026.05 | 42.77 | 36.69 | 41.62 | 24.22 | 35.56 | |
| FULLBackbone=Qwen2.5-14B and Qwen3-14B, Evaluation Protocol=Weighted average, mode=Complete memory as input2026.05 | 39.27 | 32.92 | 33.85 | 20.02 | 29.14 | |
| MemoryOSBackbone=Qwen2.5-14B and Qwen3-14B, Evaluation Protocol=Weighted average2026.05 | 38.48 | 33.02 | 37.04 | 21.04 | 30.38 | |
| MemR3Backbone=Qwen2.5-14B and Qwen3-14B, Evaluation Protocol=Weighted average2026.05 | 35.3 | 30.46 | 35.36 | 22.47 | 30.89 | |
| RAGBackbone=Qwen2.5-14B and Qwen3-14B, Evaluation Protocol=Weighted average2026.05 | 34.62 | 30.45 | 34.28 | 19.21 | 28.31 | |
| A-MemBackbone=Qwen2.5-14B and Qwen3-14B, Evaluation Protocol=Weighted average2026.05 | 32.43 | 23.3 | 33.93 | 18.48 | 23.1 | |
| MemoryBankBackbone=Qwen2.5-14B and Qwen3-14B, Evaluation Protocol=Weighted average2026.05 | 19.56 | 16.52 | 30.36 | 21.72 | 29.62 |