Databricks launches MemCurrent persistent memory for Agent Bricks agents

Databricks introduces MemCurrent, a persistent memory architecture that lets production agents carry user preferences, corrections, metric definitions and workflows across sessions.

It is built into Agent Bricks' Memory offering and already powers memory in Genie. Databricks says that on its internal benchmark, built from real enterprise data, MemCurrent triples the agent's pass rate over no memory while using 31% fewer tokens and running 26% faster. A separate Databricks blog post says it matches full-context accuracy on open-source benchmarks at about 12x fewer input tokens per question. A Databricks blog post also describes Agent Memory, a beta feature on Agent Bricks that runs on Lakebase and reports the same threefold pass-rate increase and 31% token reduction. These results are Databricks' own and have not been independently verified.