VarOps enterprise agent memory design and evaluation
What happened
On October 8, the Tessl engineering blog laid out three key design decisions in VarOps's enterprise agent memory system: storing factual statements with evidence, making time and missing-information states explicit, and deciding access permissions before calling a model. The post also covered retrieval evaluation and missing-information handling: across 500 LongMemEval questions, 99.8% had the correct evidence in the top ten retrieval results; removing the missing-information refusal turned 30 honest refusals into 30 fabricated answers. The progress reported here centers on those designs and evaluation results.
Written by AI from the coverage · updated 2 hours ago
Coverage
Follow the reports to see the story from different sides.
- AI HOT · Tips & opinionsTessl 工程博客解读企业级 Agent 记忆的三个关键设计决策
VarOps 的企业级 Agent 记忆系统采用三项设计:存储带证据的事实陈述、明确时间与信息缺失状态、在调用模型前决定访问权限。系统在 LongMemEval 的 500 个问题中,有 99.8% 的问题能在前十条检索结果中找到正确证据;取消缺失信息拒答后,30 次诚实拒答变成了 30 个编造答案。
Heat over time
Not enough continuous observations to draw a trend yet.