There is an Interpolated distribution that allows you to use samples from arbitrary distributions as a prior. Here’s the tutorial, which is badly out of date, but should work – just make sure to import pymc (NOT pymc3) and pytensor.tensor as pt (NOT theano.tensor as tt).
There’s a reason this hasn’t been updated though – it’s because it doesn’t do what you think it does. It will set the marginal posteriors as priors, but it discards all the information about the correlations between parameters in the joint posterior. The more updated approach would be to use the prior_from_idata helper in pymc-experimental. This uses a multivariate normal approximation over the entire joint posterior in the provided samples, so it preserves the dependency structure between the variables. Much preferred to the marginal-only approach.
That said, I actually endorse @Simon’s approach, with the caveat that you shouldn’t look at the posteriors one by one. If you’re only interested in reusing a subset of the prior parameters, you can compute the mean and covariance of the samples in your posterior and use that to parameterize a multivariate normal. Less fancy than using prior_from_idata, but it would get the job done.