The "trick" is to preserve the identities of the lags in the transition matrix. The corresponding block of the transition matrixIn period $t$, you have (for some state variable $x$) $$ x_t =a x_{t-1} \\ x_{t-1} = x_{t-1}\\ x_{t-2} = x_{t-2}\\ \cdots $$ So, the corresponding block of the transition matrix becomes:
$$ \begin{bmatrix} x_t \\ x_{t-1}\\ x_{t-2}\\ \vdots\\ x_{t-20}\\ \end{bmatrix} = \begin{pmatrix} a & 0 & \cdots && 0\\ 1 & 0 & \cdots && 0\\ 0 & 1 & 0 \cdots && 0\\ \vdots & & \ddots && \vdots\\ 0 & \dots & 0 & 1 & 0 \end{pmatrix} \begin{bmatrix} x_{t-1} \\ x_{t-2}\\ x_{t-3}\\ \vdots\\ x_{t-21}\\ \end{bmatrix} $$ The $a$ coefficient captures the one-period transition for $x$ from $t-1$ to $t$. The 1-coefficients carry the lags over to the next period as they are. The $a$This coefficient will be estimated and determines $x_t$, but the ones are imposed, and so aremay or may not be estimated. Note that the lag 21, which you don't want, has a zero coefficient and is not carried along.
See, for example, appendix B in Banbura, Marta and Giannone, Domenico and Reichlin, Lucrezia, Nowcasting (November 30, 2010). ECB Working Paper No. 1275