January 8, 2025
I’ve been thinking a lot about learning. Not the “read a book” kind—the real kind. The kind where you actually get better at something through experience.
A few weeks ago, I watched my nephew learn to ride a bike. He fell. Got back up. Fell again. But here’s what fascinated me: every time he fell, he adjusted. Not because someone told him exactly what to do, but because his brain was building an internal model of “how bikes work.”
Then I tested the latest AI agent everyone’s hyping up. I asked it to help with a workflow. It failed. I corrected it. Next day, same task—same exact failure.
It forgot. Or more accurately, it never learned in the first place.
The Missing Piece
This is the problem nobody’s talking about. Current AI agents are like someone who takes notes during every meeting but never reads them. They have perfect memory of the conversation, but zero understanding of what actually matters.
Here’s what I realized: AI agents don’t have beliefs.
I don’t mean beliefs like motivational poster stuff. I mean the cognitive infrastructure humans use to navigate the world.
How Humans Actually Learn
When you meet a new client, your brain doesn’t store the entire conversation word-for-word. It extracts patterns:
- “This person values directness”
- “They get frustrated when I’m late”
- “They trust me more when I show my work”
These are beliefs. They’re built from assumptions, opinions, and real experiences.
AI agents? They store everything and understand nothing.
They can’t distinguish between:
- A one-time exception
- A pattern worth remembering
- A core principle that should guide future decisions
What Needs to Change
If we want AI that actually learns, we need to stop building better chatbots and start building cognitive architectures.
The question isn’t “Can AI remember more?”
The question is “Can AI form beliefs that get stronger or weaker based on evidence?”
That’s what I’m working on. More on this soon.
How many times have you corrected the same AI agent for the same mistake?