Databricks tells that memory makes AI agents do 3 times more tasks

Databricks made agent memory available in Beta on Agent Bricks, its platform for AI agents. Memory keeps facts, preferences and workflows, and the agent does not find them again each time. Databricks tells that its internal tests show 3 times more correct results on enterprise tasks, with 31% less token use, which are units of text.

How it works

The agent keeps each conversation as a record. A background task changes records into short facts, removes copies and replaces a fact when a new conversation shows it is incorrect. At the start, the agent reads an index of the facts that it knows, and it searches for more data when it must. The memory works on Lakebase, a database that Databricks operates.

The numbers

In one test session, a question used 11 tool calls and almost 6 minutes. After memory kept the workflow, a related question used 2 calls and 41 seconds. Databricks also tells that the full conversation history used about 12 times more input tokens than good memory.

What is not known

These are tests by Databricks, and no other group made tests of them. Databricks gave no count of test tasks. The tool is in Beta, and it can change.

Insight: Memory is becoming a place where platforms compete

Databricks keeps agent memory in its database, and it tells that users can read, move and erase the memory. AgentR released a free memory tool on 9 October. This reads as a move to hold the facts that agents know about you, because kept memory makes a change of vendor hard. Test the move of memory before you select a vendor.