Zero-Mem: Zero-Token Memory Operations for LLM Agents
Summary
The paper Zero-Mem introduces zero-token memory operations for LLM agents, enabling memory without invoking an LLM or consuming memory tokens during memory tasks. It presents two complementary memory views—a entity-context graph and a temporal hierarchy—so queries can retrieve and ground evidence from retrieved traces, with only the final QA step using an LLM. The work reports significant memory-time savings (about 57.6% relative to baselines) and discusses cross-language implementation and future code release.