Databricks Releases MemCurrent Long-Term Memory Architecture, Internal Benchmark Pass Rate Tripled
Databricks has released MemCurrent, a persistent memory architecture for production Agents, combining episodic memory with derived memory and adopting three integration approaches: Plugin, Tool, and offline dreaming. Based on 114 internal test cases selected from more than 12,000 real Genie sessions, the pass rate rose from 22.3% without memory to 68.4%, while also being 1.35 times faster and 1.44 times cheaper. On the public benchmarks LoCoMo, LongMemEval, and ConvoMem, the average score was 88.1%, comparable to 87.8% with full context, but with about 12 times fewer input tokens per question (8.4K vs. 103K). The architecture has already been used in Agent Bricks Memory and Genie.
Why read this
AI score 67
The original text provides comparative data between public benchmarks and internal benchmarks, as well as ablation experiments, making clear the contribution of each part of the memory architecture to the pass rate.
Topics
View original postIntroducing MemCurrent, a persistent memory architecture designed to help production agents stop starting from zero. MemCurrent combines episodic memory with derived memory across three access patterns: • Plugins to log and load context automatically • Tools the agent calls on demand • An offline “dreaming” pipeline that consolidates memory in the background Agents carry forward user preferences, prior corrections, metric definitions and successful workflows instead of rediscovering them every session. On our internal benchmark built from real enterprise data, MemCurrent triples the agent's pass rate while using 31% fewer tokens and running 26% faster. It's built into Agent Bricks' Memory offering and already powers memory in Genie. Read the research: https://t.co/AYGXOqrhkO
View original
Introducing MemCurrent, a persistent memory architecture designed to help production agents stop starting from zero. MemCurrent combines episodic memory with derived memory across three access patterns: • Plugins to log and load context automatically • Tools the agent calls on demand • An offline “dreaming” pipeline that consolidates memory in the background Agents carry forward user preferences, prior corrections, metric definitions and successful workflows instead of rediscovering them every session. On our internal benchmark built from real enterprise data, MemCurrent triples the agent's pass rate while using 31% fewer tokens and running 26% faster. It's built into Agent Bricks' Memory offering and already powers memory in Genie. Read the research: https://t.co/AYGXOqrhkO
Body · AI translation
Databricks Releases MemCurrent: Giving Production Agents Long-Term Memory with "Online Read/Write Plus Offline Dreaming," Raising Pass Rate 3x and Saving 31% of Tokens
The @databricks team believes that the key to memory is not "storing a lot," but "what to keep, how to maintain it, and how to measure it." MemCurrent uses two types of memory plus three integration methods to simultaneously improve quality and efficiency on real enterprise tasks, breaking the "you can only choose one of effectiveness and cost" dilemma. It is already used in Agent Bricks Memory and Genie. https://t.co/QAmh3e1FWP
Three integration methods, each with its strengths · Plugin: automatically reads and writes at fixed points in a session. The advantage is predictable behavior; the drawback is that it consumes tokens even when memory is not needed. · Tool: the agent itself decides when to read and write, which is more flexible, but effectiveness is limited by the agent's own capabilities. · Offline Pipeline: batch updates in the background on a schedule or trigger condition, with low runtime cost, but updates are delayed.
MemCurrent uses all three, placing each in the most suitable position.
Two types of memory · Episodic memory: raw session trajectories, append-only and unmodified, gradually recorded by the plugin. · Derived memory: experience distilled from trajectories, stored as hierarchical files, with an automatically generated memory.md index at the top. There are two write channels: explicit user requests take effect immediately through agent tools (add, update, delete); implicit experience is distilled by the offline "dreaming" process, which is also responsible for splitting oversized files, merging duplicate content, and resolving conflicts.
Layered reading: at the start of a session, the index is injected first so the agent knows what it remembers; when needed, search_memory and fetch_memory are used for on-demand retrieval.
Key data · Public benchmarks (LoCoMo, LongMemEval, ConvoMem): average score 88.1%, comparable to 87.8% for full context, but with about 12x fewer input tokens per question (8.4K vs. 103K). On the two benchmarks with the longest histories, it even surpasses full context, because stuffing in too much content actually interferes with the model. · Internal benchmark: based on filtering more than 12,000 real Genie sessions, ultimately yielding 114 test cases. The pass rate rose from 22.3% without memory to 68.4%, closing more than 80% of the gap to the ideal upper limit of 78.6%; at the same time it was 1.35x faster and 1.44x cheaper. Mem0 performs well on public benchmarks but scores only 43.9% here. · Two examples: after one correction, the agent no longer makes the mistake of "fabricating a table without reading the query results"; after saving an exploration path as a workflow, a related new question was shortened from nearly 6 minutes to 41 seconds.
Ablation experiments show where the value comes from · Episodic memory only: 35.8%. The main problem is that the agent only thinks to search in about one-third of the cases. · Derived memory plus index: 66.1%, with most of the improvement coming from here. "Knowing what you know first" is very critical. · Combining both: 68.4%. Raw trajectories can recover details lost during distillation. · Extraction strategy: offline dreaming only 63.5%, online tools only 53.9%, combining both 66.1%.