Orthogonality
Two vectors pointing in unrelated directions — at right angles, sharing nothing.
Like you're twelve
Push a shopping trolley north while your friend pushes it east. Neither of you helps or fights the other; your efforts are independent. That is orthogonal. Now push north while your friend pushes north-east — some of their push adds to yours. That is correlation, and it is what makes stored things bleed into each other.
Precisely
Two vectors are orthogonal when their dot product is zero — cosine similarity 0, 90° apart. In practice nothing is exactly orthogonal; what matters is that random high-dimensional vectors are nearly orthogonal, with similarities hovering near zero.
Near-orthogonality is what buys interference-free storage. When you bundle several vectors, each one's contribution is recoverable because the others contribute almost nothing along its direction. When you bind, the result is near-orthogonal to both inputs, which is exactly what makes a binding act like a sealed container rather than a blend.
The failure mode is worth naming: if you encode two genuinely different things with correlated vectors, no amount of clever reading separates them again. The space is only as honest as the encoding.