PraMem turns lengthy user history into experiential memory for better behavior prediction

Large language models (LLMs) show promise for sequential behavior prediction, but they struggle with long user histories due to latent pattern induction failures and cognitive biases. Previous approaches used context-compression to reduce the historical burden, but that fails to resolve core challenges. A new paper from Chinese researchers proposes a paradigm shift: treat the lengthy historical sequence as a valuable resource instead of a burden. Their method, PraMem (Practice-derived Experiential Memory), conducts beforehand practice on the full historical sequence to build an experiential memory. This memory then serves as assisted input for accurate long-horizon behavior prediction.

PraMem's key innovation is turning passive context into active memory through practice. The model first processes the entire user history offline, learning patterns and building a structured memory representation. During inference, this memory is combined with the current context to predict the next action. Extensive experiments across diverse tasks (e.g., e-commerce, content recommendation) show PraMem consistently outperforms prior methods like memory networks and RAG-based approaches. The authors also provide in-depth analyses of how the experiential memory evolves and why it works. This work represents a significant step toward LLMs that can truly understand and predict long-term user behavior, with implications for personalized assistants, recommendation systems, and autonomous agents.