Zero-Mem: Zero-Token Memory Operations for LLM Agents

Zero-Mem is a new memory system for LLM agents that performs memory operations without consuming LLM tokens by organizing interaction traces in an entity-context graph and temporal hierarchy, retrieving relevant information deterministically without generating intermediate representations. The approach achieves competitive performance on question-answering benchmarks while reducing memory operation time by 57.6% compared to baseline methods and eliminating the computational overhead of additional LLM calls for memory management.