Unverified85% confidenceConceptTime unknown
治理Distraction型上下文腐烂的四种手段为：压缩与折叠、外部记忆（RAG）、任务切片、工具卸载
1
Sources
85%
Confidence
Long-term
Relevance
6/1/2026
First Seen
Sources
Agent上下文腐烂的两种病因：Distraction与Poisoning详解
bilibiliAI大模型原理
Related Entities
Related Claims
Unverified长上下文窗口使单Agent能在一次调用中处理复杂任务，削弱了因上下文限制而必须拆分的理由66% similarUnverifiedReAct 循环深度大时，工程界常用三类 Token 压缩策略：滑动窗口、重要性摘要、外部记忆卸载65% similarUnverified工具调用拦截层将约束从模型记忆迁移到系统硬编码，规避了上下文窗口导致的规则遗忘问题64% similarUnverifiedtokio异步运行时底层采用多线程工作窃取调度器，将大量轻量级异步任务高效分配到有限的操作系统线程上执行64% similarUnverified将复杂指令拆解为多个专注的子模块有助于减少LLM在处理长指令时的注意力稀释问题63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/22396API
curl https://kongchang.com/api/v1/knowledge/claims/22396MCP
get_claim(id=22396)