The Memory Challenge
Hosts: Chris Carolan, Casey Hawkins, Klemen Hrovat
August 18, 2026
Episode three asked why handing someone your context file still does not hand them the job, and the three hosts kept landing on the same answer: memory. Klemen Hrovat described how his three-person team stopped relying on memory altogether — one always-on machine everyone shares, and a company state that recomputes every time an email, a DM or a booking lands — while Casey Hawkins and Chris Carolan worked the other side of it, where a context file goes stale the minute you save it and your own memory of a tool can be the most out-of-date context in the room. It closed on the layers question: what belongs in shared memory, what should stay personal, and why someone else's winning prompt so often wins nothing when you run it.
Moments from this episode
Key takeaways
The honest analogy for working with AI is not junior versus senior, it is a brand-new hire versus someone who has been there for years. Experience in the craft is not the missing piece; tenure inside your environment is.
An agent does not accumulate your corrections. Every run starts on a fresh machine that gets thrown away, so a correction only survives if someone writes it back into the instructions the next run will read.
A single shared context file is outdated the minute you save it. The work is not writing the file, it is keeping it current — and that is where most teams quietly fail.
One alternative to remembering is recomputing. Klemen's team rebuilds the current state of a relationship every time an email, a DM or a booking lands, and pays for it in tokens and in a 20-30 minute wait before the draft exists.
Your own memory of a tool can be the stalest context in the room. People still refuse Breeze because of a version from years ago, and once the shared corpus says a tool is bad, most people never find out it changed.
What gets shown publicly is the hundredth invoice. Nobody posts the first 99, which is why AI looks far easier from the outside than it is from inside the work.
Memory and context are not two things. In practice both are just what you would have to tell a new coworker to get the work you actually want.
Memory has layers, and they need different homes: personal preferences stay local, company language belongs in one shared place, and shared memory needs a real process for who is allowed to change it.
Some context should stay personal for good reasons. What you most want remembered is often how a specific person wants to be handled, and writing that where anyone can read it carries its own risk.
When someone's winning prompt wins nothing for you, the prompt is rarely the problem. The memory that made it work never came with it.
Keep learning
You just watched the work. Now build it.
The next step is doing it yourself — live, with people on the same path.
One signal-dense note when a new episode lands. No noise.