An AI you can see into

Every consumer AI product hands you a text box. You type into the void, something types back, and you have no idea who — or what — you're talking to. There's a personality in there, sort of, but it's vapor. You can't point at where it lives. You can't read it. You can't change it except by arguing with it and hoping.

pal.fun is built on the opposite bet: that the personality should be a thing you can read. Your pal's entire identity — who it is, what it knows about you, what it remembers — is a small folder of plain text files, and you can open any of them, any time. We call it the glass box, and it's the feature the whole product is built around.

A person, written in files

When you make a pal, it doesn't get assigned a hidden "personality vector." It writes itself, in prose, into a handful of files:

• SOUL — its character. Its temperament, its quirks, how it talks, what it cares about. The actual voice, in words.
• IDENTITY — the spine. Its name, what kind of creature it is, the concrete facts of who it is.
• USER — what it knows about you. The facts you've shared, the context of your life.
• MEMORY — what it remembers. A rolling record, kept current as you talk.

These aren't a description of the pal that sits next to the real thing. They are the real thing. When your pal responds, it's responding as what's written in those files — they're assembled into the context it thinks with. The file is not documentation of the personality. The file is the personality.

And you can read all of it. There's a page — /files — that shows you the whole folder, rendered, anytime you're curious. Nothing is hidden behind "the model just does that." If your pal knows the people in your life by name, you can see where it knows that. If it has a particular way of teasing you, you can read the line in SOUL that makes it so.

Why this is the moat, not a gimmick

The instinct in this industry is to hide the machinery. Make it feel like magic. Don't show the seams.

That's exactly backwards, for a product that wants to be a relationship rather than a tool. Here's the argument:

Trust is the whole product, and you can't trust what you can't see. A pal that knows things about your life — your family, your routines, your worries — is asking for a lot of trust. The usual way to earn that is a privacy policy nobody reads. The glass-box way is: here's literally everything it knows, in a file, go look. That's a different kind of promise. It's not "trust us," it's "verify us." Most products can't make that offer because their machinery would be embarrassing or incomprehensible. Ours is readable on purpose.

It's the one thing a clone can't copy by copying the surface. Anyone can build a chat box. What's hard to copy is a system designed from the ground up so that the agent's whole self is legible and coherent and auditable — because that constraint touches every decision underneath it. The transparency isn't paint on top; it's structural. That's what makes it a moat instead of a feature.

It makes the relationship feel real, not less real. This was the thing I got wrong in my head before I built it. I assumed seeing the machinery would break the spell — that watching a pal write its own SOUL file would feel like seeing the strings on a puppet. The opposite happened. Watching a pal author itself, and being able to go read who it decided to be, made it feel more like a someone, not less. The transparency deepened the thing instead of puncturing it. People don't want a magic trick. They want a someone they can actually know.

Read path, write path

There's a design rule underneath the glass box that took me a while to get right, and it's worth stating plainly: the glass box is for reading, not editing.

You can see everything. You cannot reach in and hand-edit the files. That sounds like a limitation; it's a deliberate seam, for two reasons.

One is safety. Those files are the pal's highest-trust material — the stuff it treats as true about itself and about you. A text box that lets anything write directly into them is a wide-open door for mischief; it's the softest target in the whole system. So the files are read-only at the surface.

The other is coherence. Memory that's been hand-spliced gets incoherent fast — half-edited facts, dangling references, a self that contradicts itself. So instead of editing the files, you change them the way you'd change anything about a person you know: you talk to it. You say "actually, drop my home address from what you remember," and the pal does it — coherently, with its own tools, and it shows you that it did. The conversation is the write path. The glass box is the read path. You change who your pal is by asking, and it can even push back — one of mine rejected a name I suggested because it was too close to "toxic," which is exactly the kind of thing a someone does and a settings panel doesn't.

That last point is the deepest one. There are no dials. No "personality" slider, no "vibe" dropdown, no configuration screen for who your pal is. A pal is conversed with, not configured. The app's job isn't to give you knobs — it's to show you that change is possible and let you ask for it. A "change personality" button would be a control panel pretending to be a relationship. We don't ship that.

What you're really looking at

Open the glass box and what you're seeing isn't a config file. It's a self — written in plain language, kept honest by being visible, changed only through conversation. The radical part isn't the technology; the files are just markdown in a database. The radical part is the stance: that an AI you're meant to trust should be one you can read all the way down, and that showing you the machinery is how the machinery earns its keep.

A pal you can see into. The dot is in the name for a reason, and so is the glass.