The Bayesian Brain hypothesis is about how the brain makes sense of new information. And how it updates its understanding of the world based on that new information. It's involves Bayes' Theorem, which looks like this, but don't worry about it too much:
A,B =events
P(A|B)=probability of A given B is true
P(B|A)=probability of B given A is true
P(A), P(B)=the independent probabilities of A and B
To simplify, this means that the likelihood of something being true is derived by comparing new evidence to related prior beliefs. So your posterior belief depends on your prior beliefs (which are established over time) plus the new evidence that comes in.
Example: You have a prior that you’re universally hated (A), and you tell your therapist. Your therapist asks you to make a list of people who like you (B). You identify 20 people. So your posterior belief that you're unlikable lowers given the evidence you've just gathered, and that becomes your new prior - you're not as unlikable as previously thought. This is because the likelihood that people like you (P(B)) is very large - you just came up with 20 you feel pretty certain about - and the likelihood that those 20 people like you if you are in fact fundamentally unlikable (P(B|A)) is very small. So the equation then tells us that P(A|B) also reduces dramatically. Now, if your prior belief that you're unlikable is VERY strong (like in depression), it may take a LOT of evidence to really change it, but the general mechanism holds.
Priors can be conceptual, social (e.g. is it good to smile at strangers in the street? Different evidence in CA v. NYC), perceptual (e.g. faces are convex), physiological (e.g. the hot stove hurts), etc. They exist across all modalities of processing, because that's how the brain works.
Is this exactly how the brain works? I don't really care, because it's definitely effectively how the brain works. There are apparently a lot of academic debates about whether or not the brain's algorithm is exactly like that of AI systems that use Bayesian statistics. I think that's not very interesting (unless you're a computational neuroscientist). Predictive processing, which is based on Bayesian ideas, probably gets closer to the "master algorithm" (and is still very far - this is complex!) Regardless, this is a useful (and I think self-evident) conceptual understanding of how the brain holds ideas and updates them based on incoming evidence. I hope that clarifies what is meant by “priors.”