Fact Consolidation (Single-Hop) on MemoryAgentBench (MAB) FC-SH 262K
93AccuracyAblation C (gpt-4o backbone)
Evaluation Results
| Method | Links | |
|---|---|---|
| Ablation C (gpt-4o backbone)Pipeline=Ablation C (gpt-4o backbone), Backbone=gpt-4o2026.05 | 93 | |
| HeadlinePipeline=Headline, Backbone=gpt-4o-mini2026.05 | 82 | |
| Ablation A (chunk-4096)Pipeline=Ablation A (chunk-4096), Backbone=gpt-4o-mini2026.05 | 73 | |
| LLM-judgment baselinePipeline=LLM-judgment baseline, Backbone=gpt-4o-mini2026.05 | 61 | |
| GPT-4o (long-context)system=long-context2026.05 | 60 | |
| HippoRAG-v2 (best published)2026.05 | 54 | |
| BM25 (MAB chunk-512 default)2026.05 | 48 | |
| Zep / Graphiti (temporal KG)2026.05 | 7 |