让 GPT-6 Astra 长时间不跑偏的 agent 架构
作者拆解一套 Astra 智能体记忆架构：只保留当前状态、已做决策和不可重犯的失败，避免上下文被废决策塞满。
this is free f*cking gold how to make GPT-6 Astra work for hours without forgetting the goal, repeating mistakes or pretending it’s still “working” most developers give their agent more context. then it fills that context with dead decisions, failed tool calls and useless status updates. the smartest agent isn’t the one that remembers everything. it’s the one that knows what to preserve: • the current state • the decisions already made • the failures it must never repeat I broke down the exact architecture - from persistent memory to activity tracking and verifiable progress. read this before your astra agent burns another million tokens going nowhere ↓
这条是引用着另一条写的：x.com/mr_kozh/status/2097221341107925350