LiveMem: Maintaining Memory State Continuity in Long-Running LLM Inference
LiveMem keeps a separate memory state alive after older tokens leave the active context.
The paper frames this as “state continuity under context turnover” for long-running assistants and agents. LiveMem adds an intrinsic memory state to a pretrained full-attention LLM while keeping the main KV window bounded. In the authors’ LongMemEval experiments, the system could answer from memory even after supporting evidence was no longer in the current context. ArXiv · AI/CL/LG's note
The paper frames this as “state continuity under context turnover” for long-running assistants and agents. LiveMem adds an intrinsic memory state to a pretrained full-attention LLM while keeping the main KV window bounded. In the authors’ LongMemEval experiments, the system could answer from memory even after supporting evidence was no longer in the current context. ArXiv · AI/CL/LG's note
score 5