The LongMem framework enables large language models to utilize extended historical data by introducing a decoupled network architecture with memory encoding and adaptive memory retrieval components. This method significantly improves performance on long-context modeling and in-context learning tasks.