Agent Memory on Databricks

Agent Memory is a capability of Agent Bricks, Databricks' agent developer platform. It stores the facts, preferences, and workflows an agent should remember across sessions. Databricks announced it in Beta, running on Lakebase and built on its MemCurrent memory research.

What It Is

Agent Memory gives agents persistent state so they do not rediscover the same facts or repeat corrected mistakes. It has three parts: managed sessions, managed memory, and memory pipelines. Managed sessions record each conversation as an ordered history of messages, tool calls, and results in the framework's native format. Managed memory keeps profile, preferences, domain knowledge, and workflows. Memory pipelines distill sessions into long-term memory in the background.

How the memory architecture works

The design follows MemCurrent, which combines three access patterns: plugins that log and load context automatically, tools the agent calls on demand, and an offline pipeline that consolidates memory in the background. At session start the agent loads an index of what is known, then finds details through hybrid keyword and semantic search. Pipelines extract durable facts and preferences, remove duplicates, and replace facts that newer conversations contradict. Developers can give a pipeline instructions about what to retain.

Ownership and control

Sessions and memories live in a Lakebase project (Databricks' serverless Postgres) provisioned in the customer's workspace, independent of any single agent framework. Developers can read, export, and delete the entries an agent keeps about each user. Entries are partitioned by actor_id, but this is not an authorization boundary: access is enforced per memory store, and a principal with store access can read and write all actors in that store. Strict tenant or user isolation requires separate stores.

Setup path

The Agent Bricks CLI scaffolds an agent (the example uses LangGraph), declares sessions and memory in agent.toml, and deploys, which provisions the stores. A CLI command then creates a memory pipeline linking a session store to a memory store. Databricks documents a quickstart for new agents and a tutorial for adding memory to existing ones.

Vendor-reported results

Databricks says its internal benchmarks showed the architecture tripled agent pass rate on enterprise tasks while using 31% fewer tokens. In one example, a repeated question dropped from 11 tool calls and nearly six minutes to 2 tool calls and 41 seconds once the workflow was saved.

Preview costs and requirements

Current documentation states that preview users pay for the underlying Lakebase instance, with no additional Agent Memory charge; pricing may change. The AgentKit Python SDK requires Python 3.10 or newer; direct REST calls can be made from other languages. The documented put and path-list methods require databricks-agentbricks 0.6.0 or newer.