Databricks released MemCurrent, a memory architecture for agents, on October 20, 2026, alongside Agent Memory in beta on its Agent Bricks platform: enterprise agents no longer start every conversation from zero, but carry facts, preferences and working methods across sessions. On an internal benchmark built from real business conversations, adding this memory lifted completion across 114 enterprise tasks from 22.3% with no memory to 68.4%, while using 31% fewer tokens and finishing 26% faster. The model did not change at all — the entire gap comes from how memory is stored and tidied.
The trick is two layers, not more storage
MemCurrent splits memory in two. Episodic memory keeps the raw record of conversations and actions, like a ledger. Derived memory distils that ledger into preferences, definitions and workflows, like a notebook someone has already cleaned up. The split results are instructive: episodic memory alone reached only 35.8% completion, derived memory alone reached 66.1%, and both together reached 68.4%. Dumping an agent's full history into its context is far less effective than distilling it first. Across three public benchmarks, MemCurrent matched the performance of feeding the entire history while using one twelfth of the tokens, and outperformed three open-source memory systems, Mem0, PropMem and OpenClaw memory. All of these figures are Databricks' own measurements, not yet independently reproduced, and should be quoted with that caveat.
Offline "dreaming": memory is sorted after hours, not mid-task
The most distinctive design choice is an offline consolidation step Databricks likens to dreaming. While working, the agent simply writes down what happened; afterwards, a separate pipeline merges, deduplicates and resolves conflicts in the stored memory, so newer definitions override older ones and repeated facts survive only once. On the product side, Agent Memory sits on the Lakebase database: an agent starts by loading an index of what it already knows and retrieves details on demand, and Databricks says the Agent Bricks CLI can deploy a memory-equipped agent in minutes. The architecture already powers Agent Bricks Memory and Genie's data-question product inside Databricks itself.
Not the same game as consumer "remember me"
Consumer products are racing on memory too — the newly launched ChatGPT Locked Chats, for instance, keep sensitive conversations in a separate space with their own memory, which is about personal privacy boundaries. MemCurrent aims at organizational memory: a support agent distils answers from a whole team's ticket history, and the next colleague's agent inherits them. That shift raises governance questions, because a wrongly distilled memory gets reused under the banner of experience; what enterprises need to keep is the provenance chain and a way to correct entries, not just a higher completion rate. For teams building their own agents, the actionable reading is concrete: store raw records and distilled conclusions separately, give the distilling step offline time, and only then talk about a bigger model.
