Slamecka and Graf (1978) coined the term. Participants who generated word pairs (given “cold — ???”, produce “hot”) remembered them far better than participants given the complete pairs to study. The generated information carried an ownership tag — neural, not metaphorical.
The effect is robust across domains. Writing notes in your own words beats copying them verbatim. Drawing a diagram beats studying one. Explaining a concept to a peer beats listening to an explanation — even of the same concept you just explained.
Fluera’s insistence on handwriting is grounded here. Every node is generated: selected, compressed, positioned, drawn. Every Ghost Map mismatch requires you to rewrite the correction before the ideal version reveals. Copying from an LLM breaks the loop. Fluera refuses to do that.