What we're learning about maintained understanding.
Propperly is a bet that serious work should compound — that people and AI agents should build on what a project already knows instead of rebuilding it. We write to test that bet in public. Everything here is meant to be evidence-led: claims are scoped to what we can show, and articles stay in draft until they are.
Three kinds of writing — Evidence · Product thinking · Analytics practice
AI Memory Is Not a Maintained Project State
"AI memory" and a maintained project state look similar and do opposite jobs. One recalls what was said; the other tells you what is currently true.
Analytics practiceWhy Analytics Agents Need Decisions and Definitions — Not Just RAG
Retrieval finds text that mentions a metric. It cannot tell an agent which definition is current, which was ruled out, or whether the number is safe to state.
Product thinkingWhy Long-Running AI Projects Lose Their Own History
After a few months, the truth about a project lives in fragments — and neither search nor a summary reliably tells you what's still true.