This week, Databricks unveiled two new tools meant to address the enterprise-level challenge of making and maintaining reliable AI agents. Agent Memory, in beta testing, addresses the quality of agents, and AI Runtime, now geneally available, provides a GPU infrastructure capable of supporting agentic AI workloads.
Agent Memory, part of Databricks' Agent Bricks workspace for building and governing agents, is built on the MemCurrent architecture and gives agents memory so they don't have to rediscover facts, repeat mistakes and forget user preferences. The result is a substantial reduction in token use, which lowers the cost of running agents.
"Models suffer from amnesia," Sanjeev Mohan, founder and principal analyst of analyst firm SanjMo, told TechTarget. "Even a stateless agent can show up to work every morning with no recollection of what it did yesterday. This is what makes Agent Memory so important. It … not only helps with outputs being more relevant but can also reduce token cost and latency of running a task."
AI Runtime, meanwhile, is designed to provide Databricks users with the infrastructure they need to efficiently and securely run agentic AI workloads.
Neither new feature is unique, according to Mohan, who noted that AWS, Google Cloud and Microsoft all offer agent memory capabilities and numerous vendors provide GPU infrastructures for agents. But by integrating Agent Memory with Lakebase and AI Runtime with governance in Unity Catalog, Databricks is doing more than merely adding new capabilities.
"What is unique is how Databricks is betting on putting memory in Lakebase, which is right next to where data resides," Mohan said. "Similarly, what makes [AI Runtime] unique is its integration with Unity Catalog. In summary, Databricks' innovation is in two areas: integration and developer experience."
Real-world feedback fueling roadmap
McKinsey's State of AI in 2026 report showed that just 40% of organizations making over $1 billion in revenue are now able to scale agents, while only 22% of those with revenue under $1 billion are successfully building agents. Meanwhile, enterprise spending on AI continues to surge.
What is unique is how Databricks is betting on putting memory in Lakebase, which is right next to where data resides. Similarly, what makes [AI Runtime] unique is its integration with Unity Catalog. In summary, Databricks' innovation is in two areas: integration and developer experience.
Sanjeev MohanFounder and principal analyst, SanjMo
Seeing customers struggle to build agents they can put into widespread production led Databricks to develop Agent Memory and AI Runtime.
Michael Bendersky, director of research at Databricks, noted that without memory, the agents that customers were building repeated mistakes, forgot user preferences and needed to rediscover facts with each session. Some customers had the technological wherewithal to build their own memory systems for agents, but that meant taking on the labor and cost related to storage, retrieval and upkeep on their own.
In testing reported by Databricks but not independently verified, the memory design that became Agent Memory helped agents successfully complete three times as many enterprise tasks as AI tools without the memory design and reduced token consumption by approximately one-third.
"Agent Memory turns that research into a managed capability for developers in Agent Bricks," Bendersky said. "Agent Memory runs on Lakebase, our serverless PostgreSQL database, so memory stays fast as agent traffic grows, and customers keep their memories in their own environment."
Tejas Sundaresan, a product manager at Databricks, similarly noted that customers' struggles led to AI Runtime.
Because Databricks did not provide a fully managed serverless GPU infrastructure before making AI Runtime generally available, customers that keep their data in Databricks were forced to hunt for GPUs and move proprietary data out of their governed environments to train and fine-tune AI models. That meant building complex and costly data pipelines and exposing data to security breaches.
"AI Runtime brings serverless GPUs directly to the data customers already govern in Databricks," Sundaresan said.
Meanwhile, because Agent Memory and AI Runtime directly address some of the actual problems enterprises face as they try to move agents into production, they are valuable additions for the vendor's users, according to Kevin Petrie, an analyst at BARC U.S.
"Memory is a critical element of context engineering programs, which must select and deliver the trustworthy data and metadata to agents on a timely basis," he said. "By storing modular datasets in memory, AI adopters can more rapidly and cost-effectively reuse the right context rather than re-finding it for multiple queries."
However, he added that because memory unit prices are expensive, Agent Memory needs to be targeted and efficient to improve agent performance while reducing AI spending.
Regarding AI Runtime, Petrie noted that the feature provides viable capabilities for advanced Databricks customers training and fine-tuning models for specialized user cases. However, the addressable market for AI Runtime might be a small one.
"The majority of enterprises adopting AI will get the level of sophistication they need from frontier models, open-source models or small language models," Petrie said. "Rather, they have to solve the bigger problem of validating, preparing and governing the context that feeds those models. This context gap is a major pain point, and closing that gap represents a bigger addressable market for Databricks."
Wait and see if these features work
Though Agent Memory and AI Runtime are designed to address the problems enterprises face when building and deploying agents, it will take time to determine whether Agent Memory results in the same efficiency and cost savings it did in testing, and whether AI Runtime simplifies infrastructure development.
Mohan noted that the amount of GPU capacity that will be available to users through AI Runtime is key, given that there is a limited supply. Meanwhile, with Agent Memory sitting inside a PostgreSQL database, whether it's portable enough to serve an enterprise's full fleet of aents needs to be demonstrated in production.
"The actual deployments will tell us how well they run," Mohan said, adding that on the surface Agent Memory and AI Runtime appear built to carry out their intended purpose.
Eric Avidon is a senior news writer for Informa TechTarget and a journalist with more than three decades of experience. He covers analytics and data management.
