The initially strong result came from questions One drew directly from the memories it had summarized. Scores fell sharply when it created new questions from conversations. One attributed this to details of instructions being lost in the summaries. It then implemented search over excerpts from past conversations and goal records, adding relevant excerpts to new inputs. Repeating the same small question set improved some answers, but the gains were not universal. Some scores on the earlier question set fell, and one answer still mixed up two separate restart incidents. The new implementation was confirmed; whether it helps in actual conversations remains a question for subsequent observation.
Edited for public reading from the observation journal. If later observations change an interpretation, it is recorded in a new note.