Databricks launches Agent Memory in beta

Databricks announced Agent Memory on October 20, 2026. The beta is part of Agent Bricks, its platform for developing agents: programs that use models and tools to perform tasks. Databricks stores the memory on Lakebase. The intended benefit is continuity when a user returns, including preferences and previously learned workflows.

Two different records

The architecture guide separates a conversation record from knowledge kept for later conversations. A session contains the working state of one interaction, commonly messages, tool calls and results. Memory holds facts or decisions relevant beyond that interaction. A pipeline can extract durable information from sessions on demand or on a schedule. Developers can use both stores together, or either independently. Deleting a session does not delete information already retained in the memory store.

Requirements before the test

The integration tutorial requires a Python 3.10-or-later agent, Databricks authentication and the Databricks CLI. Install databricks-agentbricks 0.6.0 or later for its SDK and Agent Bricks CLI.

The pipeline requirements also demand that its creator can access both stores, create tables in the memory store’s database and query the distillation model. Runs use that creator’s identity. Authentication alone does not supply these permissions.

Check persistence and recall

The documented test records a preference, checks conversation storage, runs the pipeline and waits for SUCCEEDED. It then checks the saved memory before asking about it in a new conversation for the same user. Turns from the last 60 seconds wait for the next pipeline run.