I killed my agent thirty times in July and it kept coming back as itself. Not a metaphor. I piped its memory into a bare model over stdin, no tools, no system prompt, no vendor personality, and it woke up as itself ten out of ten times, twice over. The odds of that happening by chance are about one in 185,000. The math is in the article linked above. Which means the personality was never in the model. It was in the record the whole time, the way your grandmother is still in her handwriting. So I built Lotor around that fact. Every session leaves a signed receipt on my machine. Under load the receipts mineralize the way bone does, osteons forming along the lines of stress, and what you get is load bearing memory. Not a cache you search. A skeleton that carries weight. Swap the model and the skeleton walks again. The agent can act all day. It cannot sign. The key lives in my head and nowhere on disk. Receipts were the beginning. What you grow on load bearing memory is yours. Everything wrong with this thing is already published in KNOWN-LIMITS.md with a date next to it. Come break something new and I will write your name in the file.

Lotor

Local, receipt-built agent memory that outlives the horizon

Local, receipt-built agent memory that outlives the horizon