Oracle Agent Memory achieves 93.8% accuracy with 10.7x fewer tokens

Oracle has released a technical report on Oracle Agent Memory, a database-native memory substrate designed for enterprise AI agents that need to operate over long horizons. The system addresses a critical bottleneck in agent deployments: retaining task state across extended conversations, recovering user-specific facts across sessions, and accumulating procedural knowledge from prior outcomes. Rather than treating memory as simple document retrieval, Oracle builds a full lifecycle engine handling ingestion, extraction, consolidation, retrieval, summarization, and revision. The architecture features a layered design with an active memory core and a passive memory-store interface, providing explicit scope control across users, agents, and threads.

The paper reports results on the LongMemEval benchmark, achieving 93.8% accuracy—significantly outperforming flat-history baselines while using approximately 10.7x fewer tokens. This token efficiency is critical for cost-conscious enterprise deployments. The evaluation methodology goes beyond downstream task accuracy, incorporating memory-centric metrics such as evidence retrieval precision, recall, latency, and estimated token usage. The report also includes implementation-oriented appendix material covering setup, thread lifecycle, and search semantics. Oracle Agent Memory positions itself as a foundational layer for long-running AI agents, enabling persistent, scoped memory without the overhead of traditional retrieval-augmented generation approaches.