Cognition suggests quite the opposite
Our memory is associative, and is considered to be an emerging property of an adaptive complex system - the neural network.
Recall is reconstructive. The brain isn't a hard-drive.
You may have heard this famous quote by Wittgenstein:
The meaning of a word is its use in the language.
Similarly, with cognition:
The semantic of something is its associations.
Strip all associations from the word 'Tomato' (like something that is red, round, edible, that often goes in a salad) and it is as good as a random collection of letters like 'Buxiarw'.
What's important to remember is that neither text, nor icons are the actual semantic - they are just a step on the way to it.
What you really want people to learn is the association between the icon and its semantics. That is, create associations from a pictorial representation to some desired function.
The problem with icons
As others have mentioned, icons are ambiguous at best, or plain cryptic. To give an example, a spanner icon could mean:
So inferring the semantic of an icon is often not easy, sometimes impossible.
The problem with more
As with many aspects of life - more is less.
If an interface has only one icon, and the icon is key for (frequent) task completion, nearly anything goes. Think of the hamburger icon, that now seems so bluntly clear - there is little in the icon (when you see it for the first time) to suggest its function.
But when there are many icons, the recall ability decreases, and not in a linear way.
Putting text next to an icon seems reasonable - after all, the text clarifies the meaning of the icon.
There is a hidden assumption in the pro tip you've shared that this will suffice. But this can have an opposite effect.
Consider students who are allowed a formula sheet whilst taking an exam. Such group is far less likely to remember the formulas compared to a group that is not allowed such sheet (at least in the short term).
The brain is lazy, thus powerful
When processing information, our brain makes gross generalisations and pretty much takes any possible shortcut - doing as little as possible to get the task done.
For example, when reading, we do not read each word in its whole - we glance at the start (and end), and even this is done within a confirmatory process rather inference from scratch (see predictive brain).
This is all part of cognitive economy - within a time frame, the less needs to be done, the more can be done.
Task based behaviour
Within task based behaviour - when the user is trying to complete a task - she will normally adhere to the principle above; namely, do as little as possible.
Users don't use systems not because they want to learn them, but because they have a goal - a desired state.
So by providing a text next to an icon, the semantic becomes immediately clear and there will be little incentive to create any association at all!
The power of success
Considering a fundamental (albeit crude) cognitive modal - the TOTE model - we know that accomplishing a goal makes us feel good. Sometimes it's the release of dopamine, sometimes its the sheer reduction in cognitive load.
This is a fundamental reward mechanism.
So as counterintuitive as it may sound, letting users to struggle a bit can yield better learning and retention.
This concept is well proven within the education field: spoon-feed your students with answers and they'll learn little (à la behaviourist theories); challenge them, specifically when self-exploring, and they learn much more (à la cognitive theories).
The flow theory suggests the optimal tuning of such paradigm.
As I've argued, there is a (good) chance that if you put this pro tip to test you'll find it won't work but rather have the opposite effect. Of course, there are many variables involved so in some situations it may well work.
But more than all, I hope that the information would be transferable to other scenarios.