What compounds in thinking
Most of what an AI writes for you should be thrown away. A few small pieces make the next round of work easier.
We used to describe BaseHalf as a memory layer for coding agents. In May we sat down and asked what it was really for, and the answer was plainer: I have a folder of stuff, help me think with it. “Compound thinking” is our name for the hope behind that sentence, that what you do today makes next month’s work easier.
Saving everything does not get you there. A transcript with every word in it is about as hard to reuse as nothing at all. So which pieces do pay back?
Decisions, with the reason
Our desktop repository keeps a file of decisions, more than a hundred so far. Each says what we chose, why, and what we set aside. In September one entry recorded why we removed the file that told agents where you were looking: agents did not read it, it went stale, it got columns wrong, it cluttered git. In October we wanted something like it back, and that entry became the checklist for the new design. Without the reasons we would have rebuilt the same thing, or been afraid to touch it.
A decision without its reason is a rule nobody dares to change.
Definitions
Every line on a BaseHalf canvas means one thing: A’s context flows into B. We took labels off connections on purpose. A definition that holds everywhere means a person and an agent read the map the same way, and neither has to interpret the other’s notes on the arrows.
Your project has words like that. Write down once what “active user” or “done” means, in a note of its own, and connect it to whatever relies on it.
What you took from a source
A PDF in a folder is not knowledge yet. The note that says which two pages mattered, and why you disagree with the third, is closer. It is short, it is yours, and it can feed the next five things you write.
Questions you have not answered
An open question tells you what evidence to look for and which decision is waiting on it. Keep the question when the answer arrives: together they show how you got there.
What to let go
Drafts you rejected. Summaries that repeat the source. Clever phrases that guide nothing. Risks nobody plans to test. An agent can produce these all day, which is a good reason not to keep them. Deleting is cheap in BaseHalf, and undo brings back anything you regret.
A simple test: would this note make next month’s question easier? If so, give its card a one-line description and connect it to what it feeds. If not, let it go.