Memory as a Markov Matrix: Sample Efficient Knowledge Expansion via Token-to-Dictionary Mapping

Researchers propose modeling LLM memory as a Markov transition matrix over tokens, enabling knowledge expansion through state-space extension rather than weight updates, theoretically guaranteeing preservation of existing knowledge while adding new tokens without catastrophic forgetting.