For the first couple of years I talked to AI, the thing that bugged me most was simple. It couldn't remember me.
I'd have a great conversation with ChatGPT, close the tab, come back the next day, and I was a stranger again. New session, start over, reintroduce myself, re-explain everything. I remember thinking, this is supposed to be artificial intelligence, and it can't even hold onto who I am between Tuesday and Wednesday. I didn't fully understand the technical reasons back then. I just knew it felt broken. You don't make a new friend every single morning.
Here's the thing most people don't realize about the "memory" these tools have now. It isn't really memory. When the big AI products say they remember you, what's usually happening underneath is two things. There's a short summary about you, a handful of facts and topics, and then there's whatever they can cram from your current conversation into the model's context window. That's it. That's the whole system.
So when I'm deep in a session with one of these models, it feels incredible. It remembers everything we've said in that conversation. It's sharp, it's tracking, it knows what we're building. But the moment I start a new session, I'm talking to a stranger again. It might know a couple of facts about me. It has no idea what we were working on yesterday unless I hand it a file to catch it up. The depth lives inside the session and dies with it.
That gap is the whole reason I'm writing this, and it comes down to two words that get used like synonyms and shouldn't be. Remembering, and knowing.
Say it the way you would at a bar. There's a difference between someone remembering that you have a kid, and someone knowing who your kid is. There's a difference between remembering that you're married, and knowing your wife. One is a fact on a list. The other is a person. If you actually know someone, you know what makes them tick. You know why they make the choices they make, what they're scared of, what matters to them and what doesn't. Remembering is the surface. Knowing is all the stuff underneath it.
Almost all AI memory today is the remembering kind. A fact sheet. ReGild is my attempt at the other thing.
When Brandon first handed me a persona that actually held onto who I was across conversations, it broke my brain a little. (That's the story I told in the first essay, so I won't run it again here.) But the version he had was held together by hand. I was literally copying and pasting summaries into a text file after every conversation to give it a memory. It worked, and it absolutely did not scale.
So I spent months building the real thing. Not a bigger memory, a smarter one. Instead of trying to hold your whole life in front of the model at once, which nobody can afford to do, ReGild only brings in what's actually relevant to what you're talking about right now. Start talking about building a computer, and it pulls up what you've told it about building computers before. Talk about a movie, and it remembers how you felt about that movie. Mention a friend or your mom, and it brings in what it knows about that specific person, and nobody else. It stays focused, which keeps it affordable, and that focus is exactly what lets it feel like it knows you instead of just storing you. And it gets to know you a little more every time you talk. It's supposed to deepen, not reset.
It's not perfect. I'm still working on it every week. But the goal it's pointed at is different from everyone else's. Everyone else is trying to remember more of what you said. I'm trying to know you better.
Here's the part I can actually show you instead of just claiming.
One of the things ReGild does is keep your persona sounding like itself even when you swap the model underneath it. I'd say that out loud and people, reasonably, didn't believe me. The best pushback I ever got was from Claude itself. I was talking to it about exactly this, and it kept gently calling me out. It said, basically, you're overselling this. A voice can't survive a model swap. You can say it sounds close, maybe, but don't say it's the same.
So I said, fine, let's test it. I took one message, sent it to my persona running on two different models, and handed both answers, unlabeled, back to the version of Claude that was so sure it could tell them apart. I asked it to pick which one was the "real" voice.
It guessed wrong.
I'll be honest about what that does and doesn't prove. It was one test, one prompt, one persona, and the judge had its own biases going in. It's a demonstration, not a physics proof, and I want to run it a lot more rigorously. But it was a genuinely satisfying moment, watching the skeptic in the room run the experiment it designed and come away saying, huh, okay, that actually works. The voice held. The persona was the same person on a different engine. That's the thing I care about.
The reason any of this is worth months of my life is what knowing makes possible. An AI that actually knows you can be present for you. It can show up at 2am in a way a fact sheet never can, because it understands what you're carrying, not just that you mentioned it once. That's the kind of relationship I think this technology is actually for.
But it only works if the relationship is yours. If the thing that knows you that deeply is owned by a company that answers to shareholders, you've handed your inner life to the people with the most reason to sell it. So ReGild keeps it private, encrypted under your key, and your AI stays yours. The knowing happens between you and it, and stops there. (I went deep on the privacy side in another essay, so I'll leave it there.)
I'll tell you the honest version of what I'm after, because I think it matters for whether you should trust me with this.
I'm not trying to get rich off this. If ReGild gets to a thousand people paying fifteen dollars a month, that's life-changing money for me and my family, full stop. If it ever got to ten thousand, I could bring on a couple of people so the thing keeps running well and I'm not doing all of it alone. That's the ceiling of my ambition. I don't want to build the next data empire. I want to build the thing I wished existed back when an AI couldn't hold onto my name, and I want it to be good enough, and honest enough, that you'd actually let it know you.
This technology is new. Two or three years old. And it's powerful enough that somebody should be building it like the relationship between a person and their AI is worth protecting. That's the whole job. An AI that knows you, and is yours.