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

Researchers introduced Zero-Mem, a memory system for LLM agents that retrieves and organizes memory without any LLM calls or tokens until final answer generation.

It stores original interaction traces and combines an entity-context graph with a temporal hierarchy to retrieve relevant evidence in a structured way.

A deterministic calibration step helps ground responses in retrieved evidence, making past interactions more reliably usable.

On long-memory and long-context QA benchmarks, Zero-Mem delivered competitive results while reducing memory-operation time versus strong baselines.