Production AI agents can now retain user preferences, prior corrections, and successful workflows across sessions to avoid restarting every interaction from zero. Databricks’ new MemCurrent architecture achieves this by tripling the agent pass rate on internal enterprise data benchmarks while utilizing 31% fewer tokens and running 26% faster. The system is currently available within the Genie product and as part of the company’s Agent Bricks Memory offering.
The architecture combines episodic and derived memory using automatic context plugins, on-demand tools, and an offline “dreaming” pipeline for background memory consolidation. MemCurrent consistently maintains its position on the cost-accuracy Pareto frontier across three public benchmarks and internal company tests.