给 Astra agent 塞更多上下文反而变笨

Lummox @Lummox_eth原推发布于 2026.09.23

在 X 上打开原推

i gave my GPT-6 Astra agent more context and somehow made it worse the first version worked almost perfectly then I started “improving” it old reports went in past conversations went in examples, notes, edge cases, everything I thought might help the agent now knew 10x more about the job and started making decisions I never asked it to make it was pulling old exceptions into new tasks using outdated examples as rules and treating irrelevant history like current context so I removed almost everything instead of giving the agent the whole company, I gave it a map where the current rules live where the relevant memory lives where finished examples live and what it should never load unless the task actually needs it the next run used less context and made fewer mistakes that changed how I think about long context windows having access to 1M tokens doesn’t mean your agent should read 1M tokens good memory isn’t remembering everything it’s knowing what not to remember right now

这条是引用着另一条写的：x.com/lummox_eth/status/2100288151897981255