Pocket FM unveils Sherpa, a narrative world model that keeps long-form stories consistent

Pocket FM has introduced Sherpa, a narrative memory system designed to solve one of the hardest problems in automated long-form storytelling: keeping characters, relationships and plot lines consistent across dozens of episodes. The system is built as a Narrative World Model — a model that tracks what each character knows, how relationships shift, and which story threads remain open, then feeds that information to a language model at write time.

Instead of relying on a context window or simple vector retrieval, Sherpa constructs a dynamic knowledge graph of the fictional world. When the model needs to write a new episode, the system pulls only the facts relevant to that moment in the story: what the protagonist knows at this stage, who betrayed whom, which gun appeared in act one. Separating the retrieval engine from the writing model means the language model can be swapped without losing the narrative memory.

In an internal benchmark of 176 questions that required synthesizing information across multiple episodes, Sherpa achieved 89.8% accuracy against 57.4% for Graphiti, a competing graph-memory system. Both systems used the same language model to generate answers; the difference came entirely from the information-supply mechanism. On a public set of 576 questions the gap narrowed to 62.5% versus 51.6%, but the direction held. The tests measure detail recall, not writing quality. Whether better memory translates into better prose remains untested.

The company cites striking figures on its product page: ARR (annual recurring revenue) growing from $250 million to $500 million, a 1,200% year-over-year increase in content production, training on 5.5 billion listening hours with minute-level retention data, and 550 thousand active creators. These are vendor claims that have not been independently verified. No performance metrics on generated text quality have been published, the language model serving as the backbone has not been disclosed, and there is no discussion of inference costs for the full system.

The architectural direction Pocket FM is pushing — separating narrative world memory from the generation engine — makes structural sense and resembles advanced RAG approaches in enterprise settings. The practical question is whether the recall advantage yields a better reading experience, or merely produces stories that are equally dull but more consistent. Without an open writing benchmark — human evaluation of full prose, not trivia questions — it is difficult to assess the real value for creators and readers.