Databricks launches Agent Memory for enterprise AI

Databricks has released Agent Memory in public beta on its Agent Bricks developer platform, giving developers a native way to store facts, user preferences, and workflows. Built on the company's serverless Postgres platform Lakebase, the feature implements the MemCurrent research architecture. This design separates raw conversation histories from distilled long-term memory using a background pipeline that automatically cleans, deduplicates, and updates stored facts without interrupting active workflows.

In internal benchmarks on enterprise data tasks, integrating Agent Memory tripled task completion rates while cutting token usage by 31 percent. In one research example, an agent required 11 tool calls, eight subagents, and nearly six minutes to identify top customers for a product. With saved memory, a follow-up question required just two tool calls and 41 seconds. Databricks noted that replaying full conversation histories on open-source benchmarks consumed roughly 12 times more input tokens per question while degrading accuracy on longer contexts.

For developers and platform teams, the system eliminates the need to build custom memory infrastructure on vector databases or object storage. Managed sessions log ordered message histories in native framework formats, while managed memory uses a combination of hybrid keyword and semantic search. The memory system is integrated directly into the Agent Bricks ecosystem, working alongside the Agent Bricks CLI, Unity Gateway, Databricks Sandbox, pre-built Model Context Protocol servers, and tracing tools. Databricks claims teams can deploy a stateful agent with memory in under 10 minutes.

Practitioners retain full ownership of stored memories inside their workspace Lakebase project, enabling them to filter, export, or delete entries per user. Developers can also supply custom instructions to memory pipelines, such as specifying that an agent should save persistent customer preferences while ignoring transient debugging logs. This allows enterprise agents to maintain cross-session context, avoid repeating past errors, and eliminate duplicate context discovery during daily operations.