Ideas parked from the 1.x experimental sketch

src/agent-experimental.ts (deleted 2026-07-03) was a fully commented-out 1.x-era sketch. Nothing in it was salvageable as code, but two ideas in it are worth keeping:

1. Trajectory / experience memory. Record (prevState, event, nextState) triples per run and use them as agent experience — for evals, replay scoring, or reward-weighted decision selection. The v2 step path already produces exactly this shape (the event-sourced step envelope, see docs/p0-design.md §4.3); an experiences store is a thin adapter over it.

2. Graph-based planning to goal states. Because the machine is a graph, "plan" = a path of legal events from the current state to a goal state. A planner can enumerate candidate paths (shortest-path over the machine graph) and let the model choose among plans rather than single events — a natural extension of the decision primitive from one event to an event sequence, with the same legality guarantees at every step.

Both are post-alpha roadmap, not commitments.