ExperienceIndex paper claims up to 50.5% agent cost cuts

A new arxiv preprint proposes ExperienceIndex, a memory layer that lets AI agents accumulate and reuse knowledge about the specific documents they reason over, rather than resetting with each task. The authors report that it raises answer quality "by up to 11.0 points and reducing online dollar cost by up to 50.5%" across diverse corpora and agentic solutions.

Their framing is that current memory systems for agents mostly track "user preferences, factual attributes, or abstract reasoning patterns" and skip over what the agent has learned about individual artifacts in a shared corpus such as court cases or scientific literature. ExperienceIndex instead stores two kinds of records: "single-artifact experiences that summarize an artifact's contribution to prior tasks" and "artifact-pair experiences that encode structural relationships discovered during past reasoning." It plugs in as middleware and retrieves relevant experiences at query time to steer the agent toward the right documents. Two researchers we track flagged the preprint on the day it posted.

The paper also claims two spillover benefits: experiences gathered from text-to-SQL tasks "transfer to factoid QA tasks over the same artifact corpus," and a weaker model can "reach comparable performance" to a stronger one by inheriting its accumulated experiences. The abstract reports headline figures only, with no per-corpus or per-framework breakdown of where the 11.0-point ceiling or the 50.5% cost ceiling actually land.