EverMind(@evermind):How many memories does an answer actually need? One query needs a single fact. Another needs several pieces of evidence. A fixed top-k gives both the same budget. EverOS uses multi-round retrieval: a small model reads candidate episodes, selects the core evidence, and decides whether to stop or search again for what’s missing. In our LongMemEval-S injection ablation, core-only averaged 1.95 episodes per query: - 77% fewer answer-stage prompt tokens than k=20 - 91.0% accuracy, versus 93.4% at k=20 Retrieve again when evidence is missing. Stop when there’s enough. Let the query determine the memory budget.

2026.09.08 06:29

How many memories does an answer actually need? One query needs a single fact. Another needs several pieces of evidence. A fixed top-k gives both the same budget. EverOS uses multi-round retrieval: a small model reads candidate episodes, selects the core evidence, and decides whether to stop or search again for what’s missing. In our LongMemEval-S injection ablation, core-only averaged 1.95 episodes per query: - 77% fewer answer-stage prompt tokens than k=20 - 91.0% accuracy, versus 93.4% at k=20 Retrieve again when evidence is missing. Stop when there’s enough. Let the query determine the memory budget.

显示更多