NovelAI and the like try to work around this to some degree by having an elaborate system to inject context-specific 'memory' based on keywords (which can be re

NovelAI and the like try to work around this to some degree by having an elaborate system to inject context-specific 'memory' based on keywords (which can be recursive) and with attribute notations (genres, physical properties, etc) that their variants of the models have been trained on to minimize the amount of context it takes.