# Databricks launches Agent Memory in beta on Agent Bricks

published: 2026-10-20T17:05:26.607Z
topic: memory
source: https://www.databricks.com/blog/introducing-agent-memory-databricks
url: https://aiagentnews.io/stories/databricks-launches-agent-memory-in-beta-on-agent-bricks

Databricks has released Agent Memory in beta on Agent Bricks, its agent developer platform, to store facts, preferences and workflows that agents should retain across sessions. The feature runs on Lakebase and is built on the MemCurrent architecture from Databricks Research.

Databricks has introduced Agent Memory on Agent Bricks in beta. It stores facts, preferences and workflows that an agent should keep, and runs on Lakebase, the company's serverless Postgres. The design follows the MemCurrent architecture, which combines plugins that log and load context automatically, tools the agent calls on demand, and an offline pipeline that consolidates memory in the background.

The company says stateless agents repeat work and forget user corrections, while replaying full histories into context grows costly as they get longer. In one internal session, an agent took 11 tool calls, eight subagents and nearly six minutes to answer a question about one product's top customers. With the workflow saved to memory, a related follow-up took 2 tool calls and 41 seconds.

In internal benchmarks on real enterprise tasks, Databricks reports that the memory architecture tripled agent pass rate while using 31% fewer tokens. Sessions and memories are stored in the customer's own Lakebase project, and developers can read, export and delete the entries an agent keeps about each user. The Agent Bricks CLI can deploy an agent with memory and sessions.

## Why it matters if you build agents

Teams that would otherwise build their own memory layer on object storage or a vector database now have a managed option, though it is in beta and should be evaluated before production use. Agents can reuse saved workflows and corrections across sessions, which can cut repeated tool calls and token spend. Memory pipelines can be given instructions on what to keep, and stored entries can be audited and deleted.