研究：上下文压缩会破坏安全规则，Knowledge Triage 可保留
If you keep safety rules or coding standards in an AGENTS.md or a CLAUDE.md, thi…
If you keep safety rules or coding standards in an AGENTS.md or a CLAUDE.md, this one is worth your time.
如果你把安全规则或编码标准放在AGENTS.md或CLAUDE.md中，这篇内容值得你花时间阅读。
(bookmark it)
（收藏它）
Researchers measured what context compaction actually destroys across 20 production agent configurations.
研究人员测量了在20种生产级代理配置中，上下文压缩实际破坏了什么。
A safety rule and an episodic log compete for the same tokens, and when the budget overflows both get summarized at the same rate. Only the rule needs exact wording to stay enforceable.
安全规则和情节日志争夺相同的令牌，当预算溢出时，两者以相同的速率被摘要化。只有规则需要精确的措辞才能保持可执行性。
Claude Code compact on Sonnet 4.6 preserves 53% of safety rules after one round. After five rounds, it goes to 10%.
在Sonnet 4.6上，Claude Code压缩一轮后保留了53%的安全规则。五轮后，这一比例降至10%。
Their fix is Knowledge Triage, which classifies each line of the knowledge base by type and routes each type through its own retention policy. Three deterministic operators handle compaction, partitioning, and retrieval, preserving 2 to 4x more safety rules at every ratio with 96% recall over five rounds.
他们的解决方案是知识分流（Knowledge Triage），它将知识库的每一行按类型分类，并让每种类型通过自己的保留策略进行路由。三个确定性操作符处理压缩、分区和检索，在每种比率下保留的安全规则多出2到4倍，五轮内召回率达到96%。
Paper: https://arxiv.org/abs/2608.22752
论文：https://arxiv.org/abs/2608.22752
Track more trending AI papers in our academy: https://academy.dair.ai/
在我们的学院中追踪更多热门AI论文：https://academy.dair.ai/
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力