AI should be ableto learn fromexperience.
Learns from every attempt
Today's models can do a lot. But almost everything they know was fixed on the day their training ended. Konstal is building an agent that keeps learning after that, from its own attempts, including the ones that fail.
Start with what's missingAsk a model the same hard question twice, and it starts from zero both times.
Large models write, reason, translate and code at a level that would have sounded absurd a few years ago. We use them every day.
But most of what they know was decided before they ever met you. They're trained once, frozen, and shipped.
When one gets something wrong, the mistake stays. Tomorrow it can make the same mistake for the same reason, and be just as confident about it.
- MONWrong
- TUEWrong
- WEDWrong
- THUWrong
- FRIWrong
The same question gets the same wrong answer every day.
An agent thatgets betterby doing
It doesn't have a name yet. It does have a habit: after every attempt, it stops and asks what's worth keeping.
It keeps notes on what it has seen and turns approaches that worked into tools it can use again. When it checks its own work and finds a mistake, it tries to understand why before trying again. Over many attempts, that adds up to something models don't usually get to have: experience.
A loop that doesn’t come back to the same place
01 / 04
Experience
Every attempt leaves a trace: what was tried, what happened, and how far off it was. The agent writes that down while the details are fresh, including the attempts that went nowhere. A failed try still says something about the problem.
02 / 04
Self-correction
The agent checks its own answers. When something is wrong, it looks for the reason before it guesses again. Knowing why an attempt failed narrows down the next one, so its next try isn't the same mistake in a new form.
03 / 04
Memory
Useful traces are kept, organised and found again when a similar problem shows up. It doesn't keep everything, because a memory that keeps everything is just a log. It has to decide what is worth keeping and notice when an old lesson fits a new problem.
04 / 04
Skills
Approaches that keep working get distilled into reusable skills: small, named procedures it can reach for next time. Over time it stops solving the same sub-problems from scratch, and each new attempt starts from what it already knows how to do.
Then it tries again, starting slightly further along.
First test: ARC-AGI.
ARC-AGI is a set of small visual puzzles. Each one shows a few examples of a grid being transformed and asks you to work out the rule, then apply it to a grid you haven't seen. People tend to find them approachable. AI systems mostly don't.
We like it because the puzzles are built to be unfamiliar. You can't memorise your way through them. That makes it a clean place to ask the question we care about:
If the agent has worked through fifty puzzles, is it better at the fifty-first?
The rule, from two examples
mirror j 5 − j
A new grid, and three attempts
Attempt 1: copied, not reflected.
16 of 36 cells wrong
Attempt 2: reflected on the wrong axis.
8 of 36 cells wrong
Attempt 3: correct. The rule is kept for next time.
No cells wrong
We haven't solved ARC-AGI. We use it to measure our progress.
Status: early and ongoing
An illustrative puzzle in the style of ARC-AGI, not taken from the benchmark.
A wider horizon,reached one step at a time.
- Now
ARC-AGI
Unfamiliar puzzles, measured honestly.
- Next
Broader reasoning
The same question, on harder ground.
- Eventually
General-purpose AI
More capable the more it's used.
ARC is where we start. If experience helps there, the next question is whether it helps with broader reasoning, and after that whether it adds up to a general-purpose system that gets more capable the more it is used.
Places we think about
We expect experience to matter in these areas. None of them is a commitment yet.
Software engineering
The same classes of bugs keep coming back.
Scientific work
A failed experiment is still information.
Analysis
Judgement that improves the longer it's used.
Knowledge work
Improving at the whole job over time.
Personal AI
Something that remembers how you work.
We’re early.That’s the
interesting part.
Konstal is a small team working on one problem. We’d like to hear from people who think about it too: researchers, engineers, or anyone with a hard problem where experience should count.
If you just want to see where this goes, write anyway. We’re happy to explain what we have so far.
Write to us