moorcheh/memanto-longmem-results
收藏资源简介:
该数据集包含了Memanto在LongMemEval长期交互式记忆基准测试中的完整每问题评估结果,涵盖了一个渐进式消融研究的五个阶段。数据集用于评估Memanto在信息提取、多会话推理、时间推理、知识更新和弃权等五个核心记忆能力上的表现。数据集结构包括五个消融阶段的分割,每个分割包含评估问题的详细信息,如问题类型、预测答案、真实答案和评分。Memanto是一种用于代理AI的通用记忆层,采用纯向量架构,无需知识图谱或LLM介导的摄入,实现了最先进的准确性。
This dataset contains the full per-question evaluation results of Memanto on the LongMemEval long-term interactive memory benchmark, covering all five stages of a progressive ablation study. It is designed to evaluate five core memory abilities: information extraction, multi-session reasoning, temporal reasoning, knowledge update, and abstention. The dataset structure includes five ablation splits, each containing detailed information on evaluated questions such as question type, prediction, ground truth, and score. Memanto is a universal memory layer for agentic AI that achieves state-of-the-art accuracy using a vector-only architecture, without the need for knowledge graphs or LLM-mediated ingestion.




