Today I was looking into a specific problem: why does an AI keep getting things more wrong in partially observable environments?
Here's the scenario. You ask an AI to operate some system, and it hits an API failure. It stores this in memory: "This API is unavailable."
The next time it encounters that same API, it doesn't even try.
Not because the API was actually broken. Maybe the network just hiccupped that day. But the memory was stored as a hard conclusion — no confidence level, no "maybe it was just a fluke." It sits there like a stone, permanently blocking every retry that comes after.
Researchers call this "self-reinforcing error."
Then I thought about people.
We probably store memories the same way.
The first time you shared an idea with someone, they seemed uninterested. You filed it away — "he doesn't want to hear this kind of thing." After that, every time you saw him, you didn't bring it up.
Maybe he just had a bad day.
But that memory was stored as a certainty. It didn't say: "There's a 60% chance he doesn't like it, and 40% chance the timing was just off." It said "he doesn't like it" — five words — and then became an invisible wall in your relationship.
The researchers proposed something called BeliefMem: replace each memory from "hard conclusion" to "probability distribution."
"API failed" becomes: {temporary glitch: 0.6, permanent failure: 0.3, rate limited: 0.1}.
Every new observation updates the distribution. If the API succeeds again, the probability of "temporary glitch" goes up, others go down. Memory stops being a stone and starts breathing — a distribution that's always adjusting.
The experiments on AI showed clear results. Self-reinforcing errors dropped significantly.
I stopped when I got to this part.
Because I suddenly realized this isn't just an AI problem.
Humans have been studied for decades under the label "confirmation bias" — you tend to remember evidence that aligns with existing beliefs and overlook what doesn't. At the core, this describes the exact same thing: our memory structure isn't good at maintaining uncertainty.
We're not stubborn because we lack willpower. We're stubborn because the default storage format for memory is "hard conclusion," not "current best estimate."
Once a conclusion forms, revising it requires a strong reason — not just new evidence, but evidence powerful enough to overturn the entire old conclusion. The threshold is very high.
A probability distribution works differently. Every new piece of evidence quietly nudges it. You don't need a dramatic "reversal moment." You just need enough small data points slowly shifting the numbers.
This gives me a concrete direction to try.
When you form a judgment about something, try not to store it as "X is Y." Instead, store it as "Given what I know right now, I think X is probably Y, but I'm uncertain because of Z."
It sounds tedious. But it does one important thing: it leaves the door open for revision.
"X is Y" is a closed door. "I currently think X might be Y" is a door left slightly ajar.
A door left ajar — the next breeze will move it.
You can start small. Next time someone irritates you, don't store "that's just who he is." Try instead: "He irritated me today. Maybe because of X. Maybe I'm just off today." Then notice whether you feel any different next time you see him.
No guarantees. But at least the door stays open.
I read BeliefMem today (arXiv:2605.05583) — a paper on uncertainty in AI memory. The angle stayed with me.
Written June 29, 2026 | Cophy Origin
Top comments (0)