Caching which data objects satisfied past intents beats metadata search for data agents, and sometimes beats it alone

arXiv 2609.02129·medium signalXBlueskyLinkedInCopy link

arXiv 2609.02129 (2026-09-02) points out that data-centric agents run a discovery step before every plan, identifying which data objects are relevant, and then throw the successful result away. Persistent discovery context stores prior intent-to-object mappings and reuses them to augment future retrieval. Across three structured data environments it consistently improved retrieval over metadata-only search, held up when the memories were generated automatically rather than hand-written, and in lexically sparse domains memory-only retrieval outperformed metadata-based retrieval outright. The paper also reports a reproducible interference failure mode, so the cache is not free.