Zero-Mem cuts agent memory-operation time 57.6%

Zero-Mem removes large language model calls from every step of agent memory except the final answer, cutting memory-operation time 57.6%, per the paper posted to arXiv as 2607.29377.

That matters because memory retrieval usually bills an LLM call on every lookup, even when the agent is only checking what it already knows.

The system keeps each interaction's original trace as the record. It indexes that twice: an entity-context graph linking related mentions across conversations, and a temporal hierarchy tracking session locality. A query weighs both indexes with a deterministic calibration step. It discards evidence the two disagree on, then hands the surviving record to a single reader call. Nothing before that point touches a model.

On long-memory and long-context QA benchmarks, Zero-Mem matched the accuracy of baseline systems that do call an LLM during retrieval. The paper doesn't say what building and maintaining the dual indexes costs over a long-running agent, only that querying them skips the model.

Deterministic index-and-calibrate retrieval matched LLM-mediated retrieval here. Plenty of agent memory systems may be burning inference budget on a lookup a database could handle. Watch whether the approach holds outside long-memory and long-context QA, where most agent memory actually gets exercised.

The paper was trending on Hacker News on August 5.