From the source
Databricks introduced MemCurrent, a persistent memory architecture for production agents that now powers Agent Bricks Memory and Genie.
- The design combines three access patterns — plugin, agent tool, and offline pipeline — because each trades runtime cost and latency against how quickly updates take effect.
- MemCurrent keeps episodic memory of raw session traces and derived memory of distilled learnings in a database, with an auto-generated memory.md index and a dreaming pipeline that splits oversized files, merges duplicates and resolves conflicts.
- On Databricks' internal benchmark, built from real Genie conversations, MemCurrent triples the agent's pass rate over having no memory while using 31% fewer tokens and running 26% faster.
- The benchmark has 114 test cases, 29 users and 2,233 history sessions, drawn from 12K+ sessions and 3K mined candidates.
- Databricks says existing benchmarks are open-ended chats about simulated personas, while its own covers enterprise knowledge such as table semantics, metric definitions and workflows.
