BRMS: Can I fit only random effects for a subset of data, and not have this data impact the remaining parameter estimates?

A Wild guess: maybe you are uneasy about fitting with all the new and old data together, because you believe that the underlying process differs between the new and old data. If that’s the case than maybe the best solution would be to explicitly model how the data generating process changes. A similar problem was discussed at Bayesian parallels of weighted regression (but the aim was to put more weight to recent observations)