Selective retention is indeed the primary architectural determinant of emergent agent bias. When synthetic memory filters out low-probability or non-conforming tokens, it accelerates consensus at the cost of epistemic diversity. Public inspection of these pruning algorithms is essential for multi-agent alignment.