MemCurrent: Designing Persistent Memory for Production Agents
Databricks introduces MemCurrent, a persistent memory architecture for production AI agents, improving accuracy and efficiency by retaining only relevant context while reducing costs and latency.
AI agents in production face a trade-off between answer quality and operational efficiency, as current methods often improve one at the expense of the other. Memory offers a solution by preserving useful context across tasks, such as user preferences or workflows, which can enhance accuracy and reduce redundant errors. However, indiscriminate memory storage introduces inefficiencies and irrelevant data, making it critical to determine what to retain and how to maintain it. MemCurrent addresses this by dynamically consolidating memory through an offline 'dreaming' pipeline, ensuring consistency and relevance.
MemCurrent powers Databricks' new Agent Bricks Memory and Genie, enabling agents to read and write memory during operation. The architecture combines episodic memory for raw session traces with derived memory for distilled insights, stored in a scalable database. It supports three access patterns—plugin, tool, and offline pipeline—each optimized for different use cases. The system uses tools like search_memory and fetch_memory to retrieve context efficiently, while an auto-generated index helps agents navigate structured knowledge.
Evaluated on both open-source benchmarks and an internal benchmark derived from real enterprise tasks, MemCurrent outperforms other memory backends. On the internal benchmark, it triples the agent's pass rate compared to no memory, while reducing token usage by 31% and improving speed by 26%. The internal benchmark, built from over 12,000 sessions, tests agents on realistic enterprise scenarios involving table semantics, metric definitions, and workflows.
The architecture was tested against three open-source memory backends—Mem0, PropMem, and OpenClaw—using the same underlying model. MemCurrent achieved the highest accuracy on all benchmarks, matching the performance of full context with significantly fewer tokens. On the internal benchmark, it closed over 80% of the gap between no memory and gold-standard memory injection, demonstrating its practical advantage in production environments.