Memorable keeps the path that worked and hands it back before the next run.
The question comes back. Memorable hands the agent the sources it read last time.
The gather step replays. The compare step runs on the files it names, without re-opening the sources.
The run is stored as a chained plan, gather then compare, each step naming its output.
Claude Code and Codex sessions come in through hooks. Custom harnesses send one JSON trace. 40% fewer tool calls per task in Quartermaster on Codex.
The sources a run read, in order, become a procedure the next run follows.