What if an AI agent could manage long-term memory without spending a single LLM token on the memory itself?

This paper shows an Agent memory may not need an LLM doing extra work every time it remembers something.

Zero-Mem's memory operations used zero LLM tokens and cut latency by 57.6% versus the fastest baseline in the authors' controlled comparison.

Zero-Mem keeps the original interaction history intact, then builds two non-generative views over it: an entity-context graph for relationships and a temporal hierarchy for session and local context.

At query time, deterministic routing decides how much to rely on each view, retrieves evidence, fills in relational or nearby context, and calibrates the result before the final answer model sees it.

That means every memory operation outside final QA uses zero LLM calls and zero LLM input or output tokens.

Structured agent memory can stay faithful to raw traces without turning memory management into a recurring generation problem.

• arxiv. org/abs/2607.29377

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