Personal agents spend most of their token budget on one thing: remembering their sense of "who." (I call this Personality Tax) And "who" is mostly logistics: which apps you live in, where you journal, where your running notes sit Caching, although an obvious solution quickly becomes a foot gun if implemented naively and ends up costing more I wrote up how I cut my agent's inference bill by ~85% with multiple cache breakpoints One key idea, the stability horizon: sort the prompt by what changes, not what it means Wrote it in detail: https://t.co/XJceir0kOa