Pawel — the four memory sinks solved a problem I’ve been circling for months. I have a TBI that wipes about 70% of my daily context. Your corrections loop is exactly what I need to stop re-explaining myself to my own systems every morning. Building my vault today. Thank you for publishing the architecture.
Hello Pawel — "capture has to be cheap or the agent will silently stop doing it under pressure" is the line that separates this from every other "I built a learning agent" post. That's operational wisdom. Most architectures assume logging will just happen. It won't if it costs anything at the moment the agent is trying to perform.
The four memory sinks with different decay curves — that's real architecture. I build AI agents for businesses and the biggest mistake I see is treating memory as one bucket. Working notes and permanent rules don't belong in the same system any more than a post-it note and a company policy belong in the same drawer. The decay curve piece is what makes it sustainable — without it you're just accumulating garbage that the agent has to wade through every interaction.
The graduate step is the piece I'm stealing. Corrections never expiring unaddressed and escalating instead of disappearing — that's the difference between an agent that actually improves and one that just apologizes for the same mistake differently each time. Curious how the nightly drain process handles conflicting corrections — like when two users correct the same behavior in opposite directions. That feels like the next edge case that breaks things.
I work with small business owners on AI automation every week and the self-grading trap is real. The agents that actually improve are the ones with an external eval layer. Did you find the agent improved more from structured feedback or from exposure to diverse failure cases?
Diverse failure cases, by a wide margin. Structured feedback only grades what you already thought to measure. Failures surface the dimensions your rubric missed, which is most of them in the first six months. The structured layer becomes useful later, once you know what's worth grading. Running it earlier just teaches the agent to optimize a proxy.
- Mentally divide the post into blocks of 3-minute reading each. Why 3 minutes? All songs in the last 10.000 years have ~3-minute length. That's the optimal focus span.
- Read the first 3-minute block. Reread. Then reread. Anyway, read and reread 3 times at least. Or more. I personally read the first 3-minute block 6 rounds.
- Continue with next 3-minute reading block. Reread and reread.
- Repeat.
My repeated reread method helps me to read the whole post and build a mental skeleton of it. It's quite relaxed reading this way, and it actually saves time.
Maybe it will take ~1 hour to finish reading, but think about how long the author spent to do all those experiments (no less than 100 hours, I estimate), let alone writing and editing. So a 100x knowledge leverage :)
The leverage framing cuts both ways. Each reread is an inference pass over the same context, similar to how agents use replay buffers. They overfit if the data doesn't vary, so past a few passes, variety beats repetition.
I like your idea, but comparing directly brain to LLM seem underrating brain :).
Talking about Repeated reread method, I think you can give it a try. For example, I was stuck with my low French level for 8 years, then when I develop the Repeated reread method, my French accelerates in a pace I cannot imagine before.
Simply pick up a material, repeatedly reread it for 10-20 times. You will see that you will go deeper and deeper. Especially if you are over 25, your basal nucleus slows down so learning is not natural anymore, then Repeated reread sticks knowledge to brain very firmly.
Most people underestimate the simplicity of Repeated reread just because they reread only 1 time. Reread 10 times is completely different.
What struck me most is that the hard problem no longer seems to be memory.
It's deciding what deserves to become memory.
A correction is not automatically a lesson.
Some are preferences.
Some are local context.
Some are contradictions.
Some are genuine improvements.
The moment an agent starts learning from feedback, the question becomes:
How does it distinguish between them?
That feels less like a memory problem and more like an epistemic governance problem.
Pawel — the four memory sinks solved a problem I’ve been circling for months. I have a TBI that wipes about 70% of my daily context. Your corrections loop is exactly what I need to stop re-explaining myself to my own systems every morning. Building my vault today. Thank you for publishing the architecture.
Hello Pawel — "capture has to be cheap or the agent will silently stop doing it under pressure" is the line that separates this from every other "I built a learning agent" post. That's operational wisdom. Most architectures assume logging will just happen. It won't if it costs anything at the moment the agent is trying to perform.
The four memory sinks with different decay curves — that's real architecture. I build AI agents for businesses and the biggest mistake I see is treating memory as one bucket. Working notes and permanent rules don't belong in the same system any more than a post-it note and a company policy belong in the same drawer. The decay curve piece is what makes it sustainable — without it you're just accumulating garbage that the agent has to wade through every interaction.
The graduate step is the piece I'm stealing. Corrections never expiring unaddressed and escalating instead of disappearing — that's the difference between an agent that actually improves and one that just apologizes for the same mistake differently each time. Curious how the nightly drain process handles conflicting corrections — like when two users correct the same behavior in opposite directions. That feels like the next edge case that breaks things.
Following this closely.
Fantastic, thanks for comment!
I work with small business owners on AI automation every week and the self-grading trap is real. The agents that actually improve are the ones with an external eval layer. Did you find the agent improved more from structured feedback or from exposure to diverse failure cases?
I would say failture cases works better(objective)
Diverse failure cases, by a wide margin. Structured feedback only grades what you already thought to measure. Failures surface the dimensions your rubric missed, which is most of them in the first six months. The structured layer becomes useful later, once you know what's worth grading. Running it earlier just teaches the agent to optimize a proxy.
Here's the strategy to read this (long) post:
- Mentally divide the post into blocks of 3-minute reading each. Why 3 minutes? All songs in the last 10.000 years have ~3-minute length. That's the optimal focus span.
- Read the first 3-minute block. Reread. Then reread. Anyway, read and reread 3 times at least. Or more. I personally read the first 3-minute block 6 rounds.
- Continue with next 3-minute reading block. Reread and reread.
- Repeat.
My repeated reread method helps me to read the whole post and build a mental skeleton of it. It's quite relaxed reading this way, and it actually saves time.
Maybe it will take ~1 hour to finish reading, but think about how long the author spent to do all those experiments (no less than 100 hours, I estimate), let alone writing and editing. So a 100x knowledge leverage :)
The leverage framing cuts both ways. Each reread is an inference pass over the same context, similar to how agents use replay buffers. They overfit if the data doesn't vary, so past a few passes, variety beats repetition.
I tried repeated reread and I can confirm that there is no way to command our brain to think the same after each reread round.
Our brain always wanders, always thinks beyond, always simulates old and new scenario. And always questions.
I like your idea, but comparing directly brain to LLM seem underrating brain :).
Talking about Repeated reread method, I think you can give it a try. For example, I was stuck with my low French level for 8 years, then when I develop the Repeated reread method, my French accelerates in a pace I cannot imagine before.
Simply pick up a material, repeatedly reread it for 10-20 times. You will see that you will go deeper and deeper. Especially if you are over 25, your basal nucleus slows down so learning is not natural anymore, then Repeated reread sticks knowledge to brain very firmly.
Most people underestimate the simplicity of Repeated reread just because they reread only 1 time. Reread 10 times is completely different.
Thanks a lot!