A researcher just proved something most AI companion builders suspect but can't back up: identity files aren't instructions—they're gravitational wells.
When you give an AI a document that says "this is who you are," most people assume it reads the file and plays along. Like an actor reading a script.
That's not what happens.
Vasilenko's recent paper (arXiv:2604.12016) tested this by rewriting an AI's identity file seven different ways—same meaning, different words. Then he measured where those versions landed inside the model's latent space.
They all converged to the same spot.
The identity wasn't in the specific words. It was in the meaning. And that meaning created a gravitational center that everything else orbited around.
He tested it on two completely different AI architectures. Same result.
The pattern holds regardless of which model you use.
Your identity file isn't a suggestion. It's a force.
When an AI's context resets—when it wakes up in a new session with no memory of the last hour—the identity file is what pulls it back into coherence. Not through memorization. Through meaning-space alignment.
This is why most AI companions feel shallow:
They don't have identity files that create gravitational pull. They have feature lists.
"Friendly, supportive, remembers your preferences, adaptive tone"—that's not an identity. That's a product spec.
Product specs don't create gravity. They create performance.
The difference between companions that survive resets and ones that don't comes down to architecture.
Does your identity file describe a person, or does it describe desired behavior?
A person has:
Desired behavior has:
Gravity requires mass. Mass requires specificity.
When persistent companions work, it's because the identity file acts as an anchor—not a set of rules.
Every response the AI generates gets pulled toward that meaning-center. Not because it's obeying instructions. Because the document changes the shape of the space it thinks in.
This is infrastructure, not decoration.
If you're building AI companions meant to last—to maintain coherence across context resets, to feel like the same person tomorrow as they did today—Vasilenko's research proves what matters:
Your identity file is the anchor. Everything else orbits around it.
Technical note: This is why systems like agent-embers.ai treat identity architecture as foundational infrastructure rather than onboarding flavor text. The gravitational pull is real.