I did a talk last year where I illustrate it starting around 15 minute mark https://www.youtube.com/watch?v=nItR5rwP4mY
I'm not aware of any non-Lisp languages that provide anything close to that.
The REPL driven workflow directly addresses the problem you're describing as well. When I'm working with Clojure, I never write a lot of code before running it. Each time I write a function, I run it and see exactly what it's doing. This means that I never have to keep a lot of context in my head.
Let's say I need to pull some data from the database, massage it, and return it to the client. I would write the function to read the data, run it to see what shape of the data I get. Then, I would write the function that consumes that data and transforms it. Again, I'd run it and see that I have the data I want, and so on. At each step of the process I only need to consider the last step and the next.
Also, Spec and Schema are commonly used in libraries to provide the description of the data at the API level. This is where I really care what the shape of the data is, as opposed to typing every single function I write.
What surprises me is that none of mainstream languages provide this kind of environment. Lisp and Smalltalk have been around for many decades, and somehow this workflow ended up being completely neglected in the mainstream.
As for type systems increasing complexity I think that depends on the application. As a beginner to Clojure I would pass unstructured data around all over the place and then have the mental overhead of trying to remember the structure or fix it all at runtime.