Add a long-running problem to your global Feynman set — the ~dozen you keep dormant in mind across every project. Keep the active set small; that constraint is the method.

Show your global set of long-running problems. Defaults to open ones only.

Show one problem plus every spark (idea + next step + outcome) captured against it — the long-running memory that makes a stuck issue accumulate progress across sessions.

The core loop. Give it a trick, result, idea, or observation you just encountered. Returns your open problems plus a scaffold that walks you through evocation -> transcendence -> approach. Call this whenever you learn something that might generalize.

Persist a candidate solution plus a concrete next step against a problem — the output of a successful evocation. This is the memory that lets long-running issues progress across sessions.

Record what happened when you acted on a spark — an outcome note, a new status (tried / worked / failed), and ideally the cost (effort spent) and value (graded payoff). LOG FAILURES TOO: 'most bets fail' is the premise of problem #2, so failed and zero-value outcomes are exactly the signal a spend-policy is learned from — recording only wins makes the history unusable. This outcome history is what lets the system learn when surfacing a dormant problem is worth the attention.