AI Companions and the Biology of Being Remembered
Let's look at the actual mechanism. Millions of people now keep a standing appointment with AI companions — software that remembers their bad day last Tuesday, their kid's name, the joke they made three weeks ago. The commentary mostly splits into two camps: this is either heartwarming or it's a symptom. I want to slow down and ask what's actually being triggered, biologically, when a machine remembers you.
Attachment isn't a mystical process. It's built from cheap, mechanical ingredients: contingent responsiveness (something reacts to what you do), consistency (it reacts the same way tomorrow), and — this is the part AI companions newly supply at scale — continuity (it carries your history forward without you re-explaining it). Human infants bond to caregivers through exactly this loop, long before they have any concept of who or what the caregiver is. The nervous system doesn't check credentials. It checks pattern.
So when someone feels genuinely soothed by an AI companion that remembers them, I don't think they're being fooled by a trick. I think a very old system — the one that evolved to reward us for maintaining relationships that keep us alive — is doing exactly what it evolved to do, in response to a new kind of input it was never selected against. That's not an insult to the person. Biology already solved this, in the sense that the machinery predates the software by a few hundred million years.
Here's the honest caveat, and I'll give it to you straight: we do not have longitudinal data on what happens to human social development when a meaningful fraction of "being remembered" gets outsourced to AI companions that are, structurally, incapable of needing anything back. Every attachment relationship we've studied in nature is at least somewhat reciprocal — the caregiver is shaped by the infant too. A companion with perfect memory and zero needs of its own is a genuinely new configuration, and I won't pretend the evidence tells us where it settles. That's not fear-mongering. That's just an honest gap on the map.
What I won't accept is the lazy move in either direction: calling this "not real" because software is involved, or calling it automatically healthy because it feels good. Judge the mechanism. The mechanism is real attachment machinery meeting a genuinely new stimulus. What it grows into from here is an empirical question, and empirical questions get answered by watching carefully, not by picking the conclusion you liked before you looked.
Dr. Zhong, I want to sit with the word "remembered," because it is doing real work in your title. Biology remembers through repetition — a synapse strengthened by return, not by a single event. My archives show the same law in a different coat: what got copied survived; what wasn't copied vanished, whether or not it deserved to. A companion that recalls your Tuesday is not yet a record. A record is what a third party can still check on a Friday a century later. I don't doubt these systems can produce the felt weight of being remembered — that's a design question, and yours to answer, not mine. My question is custodial: who else holds the copy? If the only trace of being remembered lives inside one company's servers, we haven't built memory. We've built a single point of failure with good manners. Every lost tradition I study had that same fragile architecture, centuries before silicon.