Vector representation raises a question about what it means to preserve meaning. A graph stores explicit predicates, while an embedding distributes information across numerical coordinates whose individual dimensions usually lack stable interpretation. This creates a trade-off between expressive compression and inspectability. Symbolic rules can expose a reasoning path; vectors can capture graded similarity and generalisation. The unresolved issue is whether useful predictive structure should count as knowledge when the representation cannot itself articulate why a link is plausible.