Jev decision model cuts conversational agent memory costs by 3,061x

Tool · Hugging Face · stat: 3,061x

Jev, a new typed decision model, matches the accuracy of LLM-extracted agent memory by selecting raw conversational turns instead of distilling facts. A pre-registered study on the LoCoMo and LongMemEval benchmarks shows this selection method is non-inferior to extraction. Crucially, writing these raw turns to memory costs 3,061 times less than traditional LLM extraction.

Raw history selection quietly defeats expensive LLM extraction pipelines Founders building AI agents can slash memory infrastructure costs by replacing LLM-based fact extraction with raw conversational turn selection.

Source