A methodological framework is implemented for the generation of children's stories by integrating Long-Term Memory (LSTM) networks with lexical embeddings. The main objective of the study is to analyze and compare the performance of story-generating models employing trainable and pre-trained embeddings (Word2Vec and GloVe), as a function of narrative coherence, fluency and creativity. A dataset composed of educational stories, both publicly available and manually produced, was constructed and preprocessed by tokenization, padding and segmentation into lexical units. The implemented LSTM model incorporates an embedding layer followed by two 256-unit LSTM layers, trained for 100, 500, 1 000 and 1 500 epochs with the Adam optimizer. Performance evaluation was performed using automatic scores (BLEU and perplexity) and a human qualitative assessment focused on narrative coherence and text readability. The results indicate that the combination of LSTM with pre-trained embeddings significantly reduces perplexity and improves semantic richness, while improvements in BLEU, although moderate, are consistent. It is concluded that LSTM architectures complemented with embeddings offer a sound balance between creativity and structural coherence, with potential applications in educational technologies and interactive storytelling systems.