I used to think this but then improved and found that it was a matter of tooling. Using beads[1] and getting the AI to write its own tickets and wiki articles using GitHub's wiki feature on repos, it can keep track of long horizon projects pretty easily. Then you curate some of the tickets or read through them and have them update the beads or adjust them as they work them. When they want to retain some important piece of knowledge you have them add it to the GitHub repo wiki.

Modern AI programming looks an awful lot like traditional product owner or product management roles.

1: https://beads.gascity.com/

Then you're really really bad at using it. At the very least it is functional with minimal effort on codebases with 10s of thounsand