Every morning I open a fresh session and the agents know nothing.
Not a little rusty. Nothing. The session is the open conversation I have with the AI tool I build software with. Yesterday we settled a decision, backed out of a dead end, learned why one approach would never work. Today none of that is there. The agents start clean.
So I re-explain. I tell them the thing we already decided. I watch them walk straight into the dead end we left last week, confident, like it is brand new. I correct it. Tomorrow I will correct it again.
It is expensive. Not just my time. Every lesson costs something to learn the first time, and I am paying for the same ones over and over.
So I ask the agents to build a memory that survives. A place where the decisions we make and the lessons we learn get written down, and pulled back up when the next session needs them. The agents do the writing and the recall. I decide what is worth keeping. Nothing about the model itself changes — it is not learning in the way training would. It is just that the notes from yesterday are in the room today, instead of lost overnight.
Then I test it. New session. I say nothing about yesterday. And the context is already there. The decision still holds. The dead end is marked as a dead end. The agents start from what we know instead of from zero.
Learnings
Without a memory, you re-explain everything and re-pay for every lesson. I was treating each morning as a fresh start, and it was quietly costing me the same corrections again and again. Once the decisions and the dead ends carry over, the work compounds. The next session is not starting over. It is starting from everything the last one already figured out.