PraMem reframes historical sequences as experiential memory for long-horizon behavior prediction

PraMem is a new approach to long-horizon behavior prediction that addresses the limitations of large language models in inducing latent behavioral patterns and overcoming cognitive biases. Instead of compressing historical data, it conducts beforehand practice over lengthy sequences to build an experiential memory used as assisted input.

• Reframes historical sequences from a burden into a valuable resource for exploitation.
• Builds experiential memory through beforehand practice on the historical sequence.
• Uses this memory as assisted input to improve prediction accuracy.
• Demonstrates superior performance compared to prior methods across diverse tasks.

The authors consider this significant because it resolves core challenges in long-horizon prediction by leveraging history rather than just alleviating its burden.