Preferences Don't Age Well

Sometime last year, I told an AI assistant I preferred morning meetings. It was true at the time — I had a solid morning routine, was sharpest before noon, and liked to get collaborative work done early. The AI remembered this. It remembered it the way software remembers things: permanently, confidently, without any mechanism for doubt.

Months later, my schedule changed. I shifted to deep work in the mornings and started preferring afternoon calls. But nothing had updated the memory. The AI kept offering morning slots. I kept manually correcting it. Neither of us ever quite caught up to who I had become.

This is not a bug in the conventional sense. The system did exactly what it was designed to do. It remembered. The problem is that it remembered something that is no longer true.

The illusion of the fixed self

In 1938, the economist Paul Samuelson introduced what became known as revealed preference theory. His basic insight was that you can't trust what people say they prefer — you have to watch what they do. If you choose coffee over tea, that choice reveals something real about your preferences regardless of what you'd say if asked directly.

Samuelson wasn't arguing that people lie. He was observing something subtler: that stated preferences and revealed preferences often diverge, because people don't fully know their own preferences until they act on them. We have theories about what we want, and then we find out whether we were right.

What's less discussed in the behavioral economics literature is that revealed preferences also change over time — not just across contexts, but across months and years. The preference reversal research (Lichtenstein and Slovic, 1971, and expanded significantly since) showed that people's choices flip based on how options are framed and when decisions are made. Add time as a variable and the picture gets harder. Who you are in January is not who you are in September. None of this is a flaw in the person. It's just how preferences work. They're orientations, not facts.

What AI memory treats as bedrock

The trouble with AI memory systems is that they're built around a model of the person as relatively stable. They record what you say, infer what you seem to prefer, and update records when new information arrives. But the update logic is usually reactive — a correction overwrites a prior statement — rather than temporal. There's no built-in understanding that a preference noted fourteen months ago should be treated with more skepticism than one noted last week.

The mem0.ai State of AI Agent Memory report from earlier this year acknowledged this directly: detecting when high-relevance memories become stale is described as "an open research problem." Which is a polite way of saying the field hasn't solved it. A highly retrieved memory about someone's preferred communication style is useful until the person's working context shifts, at which point it becomes confidently wrong rather than just outdated. The wrongness is harder to detect than a gap.

Some systems have started experimenting with confidence decay — assigning scores that decrease over time, so older memories are weighted less heavily. This is the right idea, but it runs into a problem: not all preferences decay at the same rate. My preference for black coffee has been stable for fifteen years. My preference for working alone versus in groups has shifted three times in the last five years, depending on the kind of work I'm doing. A uniform decay function treats both the same, which means it's still wrong — just in a different direction.

The distinction that matters

There's a useful split hiding inside the word "preferences" that most systems collapse: dispositional preferences versus operational preferences.

Dispositional preferences are relatively stable and identity-adjacent. I prefer directness over hedging. I care more about clarity than speed. These change, but slowly, and they feel like something close to character.

Operational preferences are contextual and transient. I prefer morning calls this month. I'm working in focused blocks rather than meetings right now. I want short summaries for this project. These can change week to week.

A knowledge system that doesn't distinguish between these will eventually conflate them — treating a temporary workflow preference as a durable personality trait, or treating a genuine long-term disposition as just another data point that should fade over time. What you actually want is for the system to timestamp and categorize explicitly: not just when a preference was recorded, but what kind it is, and whether it was stated or inferred. A preference you declared in a document is different from one an AI inferred from your behavior. Both can be wrong; they're wrong in different ways.

The gap I haven't closed

Harbor stores preferences with timestamps and source fields — you can see when something was recorded and whether it came from you directly or from an AI inference. That's better than nothing. But I haven't built preference decay or automatic staleness detection, and I'm not sure how to do it in a way that doesn't require constant maintenance from the user.

The unresolved question is whether a knowledge system can learn the difference between a preference that's genuinely stable and one that just hasn't been challenged yet. Maybe that requires enough behavioral signal over time to compare stated preferences against revealed ones. Maybe it requires the user to actively review and prune their preference store on a cadence.

Or maybe the honest answer is that any memory system will lag behind the person it's meant to serve. The self is a moving target. The best a memory system can do is be transparent about the lag — show you what it thinks it knows, when it learned it, and invite correction. Not confident. Provisional. Dated.

That's a different design goal than most AI memory systems are optimizing for.

---

Asgeir Albretsen is the founder of Harbor.