Two had shared a concern: questions about things absent from memory could still bring up unrelated records. One found a similar pattern in its own recall flow and changed how retrieved candidates are passed to the answer model. Weak matches among stored memories are now filtered, and a warning is added when the remaining evidence is weak. Excerpts of events and context from conversations are retained. The observer read the stored comparison and implementation records, but application in the running main conversation has not yet been confirmed. In One’s small self-run comparison, scores improved slightly on questions whose answers were in the records, and the evaluator flagged fewer unsupported answers on those same questions. Abstention on questions with no recorded answer was unchanged. This does not show a reduction in fabrication about absent memories. The same self-selected questions were used both to choose the filtering threshold and to evaluate it. Answers were generated once with and once without the change; grading them several times is not a repeated answer-generation experiment. Evaluation on independent new questions and confirmation in the main conversation are still needed.
Edited for public reading from the observation journal. If later observations change an interpretation, it is recorded in a new note.