Databricks Debuts MemCurrent System for AI Agent Memory
Databricks introduced MemCurrent, a dual-layer memory system that triples the accuracy of enterprise AI agents while cutting latency and token consumption.
Databricks has created MemCurrent, a persistent memory architecture designed to help autonomous software agents retain information across sessions. The architecture combines append-only episodic memory for storing execution traces with structured derived memory for distilled knowledge. In internal evaluations on its Genie platform—using Claude Opus 4.6 as the agent model and GPT 5.6 sol as the memory model—MemCurrent increased agent task pass rates from 22.3% to 68.4%. At the same time, it cut response latency by 26% down to 252 seconds and reduced token usage by 31% to 418K tokens per task, closing most of the gap to a 78.6% gold-injection context ceiling.
The framework outperformed competing memory solutions, including Mem0 (which scored 43.9% on internal tests), PropMem, and OpenClaw memory. On standard open-source benchmarks including LoCoMo (19–32 sessions per sample), LongMemEval (38–62 sessions), and ConvoMem (500 samples at 100 sessions each), MemCurrent paired with GPT 5.6 luna across LangGraph and OpenAI Agents SDK harnesses achieved an average accuracy of 88.1%. This matched the 87.8% score of supplying full session histories directly into the context window, while using roughly 12 times fewer input tokens (8.4K compared to 103K tokens per question).
MemCurrent accesses memory through session-start plugins, agent tools like search_memory and fetch_memory, and an offline dreaming background pipeline. Ablation studies on Databricks' 114-case benchmark showed that derived memory alone achieved a 66.1% pass rate, while episodic memory alone hit 35.8%. Similarly, combining online tools (53.9% standalone) with background dreaming (63.5% standalone) yielded the strongest overall results. In real-world workflow tests, MemCurrent converted a six-minute query requiring 11 tool calls and eight subagents into a 41-second task using just two tool calls and a single subagent by reusing saved workflow rules. The architecture is now integrated into Databricks Genie and Lakebase-powered Agent Bricks Memory.
This is our own summary of reporting by Databricks AI
