By Dr Raj Kadiwar and Krishna Kadiwar.
Standing Questions: a memory pattern for long-horizon AI-agent projects, distilled from ten weeks and a few hundred sessions of getting agent memory wrong. Code, schemas, and metrics: github.com/Rocco-alt/standing-questions. Canonical home of this post: standingquestions.com.
A note on epistemics before anything else: everything below comes from the two of us running one production project. Where we state a fact we tag it [observed] (we saw it in our own logs and artifacts), [inferred] (our interpretation of what we saw), or [literature] (someone else's published result). There is no control group. Nothing here is a benchmark. It's a field report with the receipts we can show.
The incident
For ten weeks we ran a production ML project almost entirely through a coding agent - Claude Code, working the same repository session after session; roughly 300 working sessions by the end, counting the continuation lanes long sessions split into [observed].
