Knowledge Objects: Hash-Addressed Facts Hit 100% Accuracy at 252x Lower Cost Than In-Context

ArXiv paper 2603.17683 benchmarks a deceptively simple idea: instead of stuffing facts into your agent's context window and hoping the model remembers them, treat each fact as a discrete, hash-addressed tuple stored externally and retrieved on demand. The results are stark: 100% accuracy across 7,000+ facts versus in-context approaches where compaction loss destroys 60% of facts in production systems. The cost difference at scale is 252x. arXiv

The architecture treats facts as first-class addressable objects — each gets a content hash, typed metadata, and a retrieval interface. This is the database approach applied to LLM memory: instead of a growing context window that degrades as it fills, you get a stable key-value store where retrieval precision doesn't decay with volume. The 252x cost reduction comes from the obvious place: you stop paying to process 7,000 facts on every inference call and instead retrieve only the relevant subset.

For anyone building agents with persistent memory — and that's increasingly everyone — this paper provides the architectural pattern that actually works. The in-context approach that most production systems use today is a known failure mode being tolerated because alternatives weren't benchmarked. Now they are. Hash-addressed knowledge objects are the RAG equivalent of moving from flat files to a database: same data, fundamentally different reliability characteristics.