Two fractional orders for the memory of meaning: the diffusive representation of the semantic trajectory of discourse
Year: 2026
Type: preprint
Abstract: Written language has a long memory: what a passage is about still influences the texthundreds or thousands of words later, and the influence fades slowly, as a power of the distancerather than exponentially. This matters for understanding how discourse is organised and forartificial intelligence, since the language models that process long texts must decide how tostore the past. Fractional calculus describes such a memory as the combined effect of manysimple memories, each fading exponentially at its own speed. Replacing every word with itsword vector, a list of numbers that encodes its meaning, turns a text into a path through aspace of meanings whose memory can be measured. We analyse novels in English and Italian,the debates of the British House of Lords and philosophical and scientific treatises. In allof them one fractional law is not enough: the memory is the sum of two, a fast one actingfrom word to word and a slow one that takes over beyond about a hundred words. The twolaws predict an independent measurement of persistence, survive a change of language and ofword vectors, leave nothing in the dynamics unexplained and carry different vocabularies.What sets narrative apart is the slow memory: in novels it has a characteristic length of a fewhundred words, the size of a scene, while in debates and treatises it has none. The result is acompact, measured description of how texts remember, and a prescription for the memory oflanguage models, which today assume a single law chosen in advance.
Author: Fazzino, Stefano
Language: English
Subfield: Modeling and Simulation
Field: Mathematics
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7234561187
