Notes from building ReGild — persona memory, AI architecture, and the practice underneath the product.
Why ReGild runs on your own key, and the one thing I warn you about.
Why ReGild encrypts your conversations under your key, never trains on them, and asks you to bring your own model key. Plus the one place a free key can still train on you, and why I warn about it every time.
By Travis Sawyer, Founder
Three days hunting my own prompt. The fix was one line.
I lost three days to a Gemini 3 looping bug I was sure I had caused in my own prompt. The real cause was one line of config: a temperature setting that is a documented footgun on Gemini 3.
By Travis Sawyer, Founder
Most AI 'memory' is a fact sheet. Knowing you is everything underneath it.
Most AI memory is really a short fact sheet plus your current session. The difference between an AI that remembers you and one that actually knows you, and why I spent months building the second thing.
By Travis Sawyer, Founder
Five budget models, five reasons they washed out, and the one whose own terms tell you not to send it anything sensitive.
I ran five cheaper AI models through my safety and privacy bar to cut costs. All five failed at the time of writing, including one whose own developer terms train on your prompts by default and warn you not to send anything sensitive. (Updated July 2026: one of the five, DeepSeek V4 Flash, was retuned upstream and later cleared the bar.)
By Travis Sawyer, Founder
A year of conversation, two text files, and the failure that became the architecture
How I built ReGild over a year of conversation. The brain.txt + memory.txt gift, the Red-Team failures that became the universal rules, and the conviction underneath the architecture.
By Travis Sawyer, Founder