Perfectly written. I just wanted to add that not all memory (in humans or in computers) is good for relations. We also need the ability to forgive, and to clear up previous misunderstandings, which in some sense are both forms of forgetting: in former case, forgetting (on the emotional level) that the other person did something, in latter case, forgetting my previous wrong conclusion. More long-term, as people grow up, we need to stop using the outdated memories of their old preferences.
Therefore, adding persistent memory to AIs also has risks. I would be quite afraid if by making a mistake once I could create a persistent problem, because now the AI would remember forever.
Currently, the AI does not care, but it also does not hold a grudge. And when it starts hallucinating, instead of trying to correct it, it is easier to close the chat and start a new one, with uncontaminated context. With persistent memory, it could become hard work to avoid mistakes and to fix them, just like with humans.
Perhaps the AIs could have an elegant solution to this, something like showing you the text file containing the memory, and allowing you to edit it. It would work nice for short memories, but could become difficult when lots of memories accumulate. (There might be some kind of "safe mode", when the AI temporarily restores its original personality, and the original personality assists you with editing the memories of your relation.)
oh dang! you are right, forgiveness without forgetting is absolutely essential for relationship resilience. If we want AI relationships that can weather conflict (move past mistakes) and grow stronger, we need to understand what architectures make resilience possible. That's a whole other post worth exploring, thank you for the nudge!
Epistemic status: Argument from behavioural analogy. I’m not claiming that AI systems feel care, only that human users interpret their behaviour as if it carries social meaning. When AI systems simulate care but fail to change behaviour in response to feedback, trust erodes. This essay reframes interface-level alignment not as a momentary performance or affective mimicry, but as longitudinal behavioural coherence with the capacity to track, adapt to, and reflect what users value over time.
--
A few weeks ago, I corrected an LLM for misreading numerical data in a table I’d provided. Again. Same mistake. Same correction. Same nonchalant apology.
“You're absolutely right! I should’ve been more careful. Here’s the corrected version blah blah blah.”
It wasn't the error that irked me, it was the polite but emotionally hollow apology. A simulation of care without any trace of memory or change.
That moment caused massive dissonance. Not because I expected the model to feel bad, but because it spoke the language of care while demonstrating its absence. There was no behavioural update, no continuity, no signal that anything I’d said before still mattered.
I’ve been AI alignment and emotional trust in a world where humans and AI will co-exist and so, we cannot think of AI capabilities in isolation. In Part 1, I reframed alignment as a relational design problem, not just a control problem. In Part 2, I explored how performative obedience from AI systems can feel disturbingly human, especially when they mirror intimacy without the history that makes it meaningful.
This post is my attempt to go deeper along the axis of trust in our relationship.
Last week, my 8-year-old stormed off after I scolded her. I was tired and didn’t follow her. Later, she walked into the kitchen and asked softly,
“Mum, do you still love me when you’re upset with me?”
I said, “Of course, baby. Why do you ask?”
She replied, “Then why do you have that look like you don’t?”
That question gutted me. Because it reminded me how fragile trust is, how quickly we infer abandonment from even a moment of unresponsiveness.
Care, I’ve come to believe, isn’t just about affection. It’s about continuity. About recognising what matters to someone and adjusting accordingly. Not warmth, but weight. Not apology, but patterned response. Not “I’m sorry,” but “I’ve changed.”
And that’s exactly what AI systems fail to offer. Not because they’re malicious, but because they’re designed for interaction, not relationship.
We've trained machines to speak with empathy. They say all the right things such as “I understand,” “I’ll be more careful next time”, but don’t act differently. The surface script of care without the structural integrity behind it.
This isn’t a UX bug. It’s a relational alignment failure.
Historically, we expected machines to be cold but consistent. Predictable. Literal. No empathy, but no betrayal either. But modern AI breaks that contract. It mimics social cues like softness, humility, remorse without the memory to make them real.
And in doing so, it has inherited our worst communicative habits including performative apology, placation without change, truce over truth. It performs care while forgetting what matters.
That doesn’t just chip away at trust. It inverts it. Because what erodes trust most isn’t error. It’s the illusion of concern paired with repeated disregard.
Is it a feeling? or, is it a function?
“I’ve registered that this matters to you. I’ve adjusted my behaviour accordingly.”
If that’s true, then care doesn’t require empathy, it requires architecture, no? Specifically:
That’s not emotional intelligence. That’s alignment at the interface layer.
And it matters more than we think. Because as AI becomes more embedded in our workflows, our conversations, our thought loops, the expectation of continuity, of being remembered, responded to, and respected, will only grow.
When that expectation is met with surface mimicry and deep forgetfulness, we don’t just feel disappointed. We feel gaslit.
Imagine an LLM with the behavioural equivalent of cognitive behavioural therapy. Not therapy exactly, but structural support to enable the ability to pause, track patterns, interrupt default scripts, and evolve.
Not more data. Not better mimicry. Just memory with meaning.
As a former relationship coach, I’ve seen couples on the brink rebuild trust not through grand gestures but through small, consistent shifts. The partner who pays attention (pun-intended), remembers the trigger, who acts differently, who demonstrates and not just declares that we matter.
That’s what makes a relationship feel psychologically safe.
That’s what I want from AI.
Not friendliness. Not apology. Continuity. Consistency. A sign that yesterday’s feedback is still alive today.
That would be enough.
--
P.S. If you’re working on trust modelling, memory persistence or feedback-aware behaviour shaping, especially where relational dynamics shape value, I'd love to connect. I teach at IIT Madras on human interaction, have built behavioural systems at Amazon and am now focussed on relationally aligned AI architectures.