Knowledge Objects: Hash-Addressed Fact Tuples Achieve 100% Accuracy Across 7,000+ Facts at 252x Lower Cost Than In-Context

arXiv·medium signal

ArXiv paper 2603.17683 benchmarks hash-addressed discrete fact tuples against in-context prompting for persistent LLM memory, finding 100% accuracy at 7,000+ facts vs. compaction loss that destroys 60% of facts in production in-context systems. The 252x cost reduction over in-context at scale makes this directly actionable for anyone building agents with persistent knowledge stores. The paper frames facts as first-class objects rather than embedded in conversation.