Excellent classic review paper from Walter Kintsch.
The main thesis: interventions which help people remember texts they read are not necessarily those which will best help people learn from what they read. (see: Retrieval practice and transfer learning!)
He explains this by way of his construction-integration model: learning requires using a text to construct a situation model encompassing both the text’s propositions and prior knowledge. This is not the same as being able to remember the text: one could do that with only on episodic memory (albeit with limits). But with a well-connected situation model, elements of the text can be retrieved through various routes, permitting both increased recall and more paths for inference. (Spaced repetition memory prompts should encode ideas from multiple angles
For example, he cites a study which compared advance organizers (Ausubel, D. P. (1960). The use of advance organizers in the learning and retention of meaningful verbal material. Journal of Educational Psychology, 51(5), 267–272) organized so as to be consistent or inconsistent with a target text. Consistent organizers improve memory performance better, but inconsistent organizers produce much more improvement on inference tasks (i.e. learning).
He also suggests that “the best text form” will depend on a reader’s prior knowledge, citing a study which found that low-knowledge readers learn best from high-coherence texts, but the opposite is true for high-knowledge readers. This might be because “the text was so easy for them that they felt they understood it well, without actually being sufficiently challenged to work out all the details. A feeling of understanding at the level of the textbase can conceal incomplete understanding at the level of the situation model.”
He closes by pointing out that “the problem for learning from text is that reading can be too easy: it takes very little effort to read most texts with a good sense of understanding. All too often, however, that understanding is only superficial.”