MegaMem preprint reports improved retrieval across 650M tokens of enterprise records

Researchers report that MegaMem, a system for retrieving evidence for language models, improved the Overall score on EnterpriseRAG-Bench from 68.22 to 82.26 and reached 86.50 Correctness. The benchmark contains more than 500,000 enterprise documents and approximately 650M tokens. MegaMem searches condensed records and detailed evidence, links condensed matches to original sources, and reranks the results before passing selected evidence to the answer model. This lets it search a larger memory while keeping the material used to generate answers within a fixed budget. The preprint's measured results cover the benchmark; its proposed reach to one billion tokens extends beyond that evaluation.