Databricks MemCurrent Improves Agent Pass Rate 3x and Cuts Token Usage by 31%

• Databricks introduced MemCurrent, a persistent memory architecture for production agents, currently applied in Agent Bricks Memory and Genie.
• Internal benchmarks show a 3x improvement in agent pass rate (22.3% to 68.4%) compared to no-memory baselines, alongside a 31% reduction in token usage and 26% faster execution.
• The system combines three memory access methods (Plugin, Tools, Offline Pipeline) and two memory types (Episodic and Derived) to balance quality and efficiency.
• On open-source benchmarks (LoCoMo, LongMemEval, ConvoMem), MemCurrent achieved 88.1% accuracy using approximately 12x fewer input tokens (8.4K vs 103K) than full context injection.

Architecture and Design

• Episodic memory stores raw session traces via append-only Plugin writes, while Derived memory extracts structured knowledge into files indexed by memory.md.
• Explicit user requests are handled by Agent tools (add_memory, update_memory), whereas implicit learning is processed by an offline dreaming pipeline that consolidates and deduplicates data.
• Reading occurs via Plugin injection of the memory index at session start and search_memory or fetch_memory tools during execution.

Evaluation and Results

• The Internal Memory Benchmark consists of 114 test cases from 2,233 history sessions across 29 users, focusing on enterprise data workspace tasks.
• MemCurrent outperformed Mem0 (43.9% pass rate) and PropMem, closing over 80% of the gap between no-memory and gold injection ceilings.
• Ablation studies indicate Derived memory contributes most to performance (66.1% pass rate alone), while the combination of offline dreaming and online tools yields the best extraction strategy (66.1% pass rate).

Real-World Impact

• In practice, MemCurrent prevented hallucinations by storing past SQL execution errors and reduced computation costs by caching workflows, cutting tool calls from 11 to 2 for similar queries.
• The memory system acts as a portable knowledge base, allowing users to migrate between agents or LLMs without losing organizational context.