Introducing Agent Memory on Databricks

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Introducing Agent Memory on Databricks · Channel: Databricks · 01:04 · Watch on YouTube

Topics: AI agents, Data, databases and embeddings, Prompt and context engineering

Summary

Agent Bricks adds agent memory, built on Lake Base, to help agents retain useful information between conversations. Its three components are managed sessions, managed memory, and memory pipelines, which distill session history into accessible memories. The Databricks research team found that agents with memory passed three times as many tasks while using 31% fewer tokens. Users can connect session and memory stores to create a pipeline, and the feature is now in beta on Agent Bricks.

Key points

Agent memory is built into Agent Bricks, with all three components built on Lake Base.
Managed sessions record each conversation turn by turn.
Managed memory stores useful information about users, including preferences, facts, and workflows.
Memory pipelines connect sessions and memory.
Pipelines distill sessions into memory, merge duplicates, and keep memory files easy to access.
Databricks research found that agents with memory passed three times as many tasks using 31% fewer tokens.
Setup involves binding a session store and a memory store, then creating a memory pipeline between them.
Databricks automatically synthesizes new memories in the background.
Agent memory is in beta on Agent Bricks.

Chapters

Start	Chapter	Summary
00:01	Why agents need memory	The video introduces the problem of agents forgetting previous conversations and presents Agent Bricks memory.
00:33	Memory pipelines and results	Pipelines turn session history into accessible memory, and research reports improved task performance with fewer tokens.
01:04	Agent memory beta	The video closes by describing how agents can use what they learned from earlier conversations and announcing the beta.

Quotes

Most agents start every conversation from zero. — 00:01

The Databricks research team found that agents with memory passed three times as many tasks, using 31% fewer tokens. — 00:33

Bind a session store and a memory store, then create a memory pipeline between the stores. — 00:33

Your agent’s next conversation starts with what it learned from the last one. — 01:04

Concepts and entities

Agent Bricks, Lakebase, Databricks, agent memory, managed sessions, managed memory, memory pipelines

Full transcript (original language)

00:01 Most agents start every conversation from zero. >> [music] >> They forget what users told them and redo work they’ve already done. >> [music] >> That’s why we built agent memory into Agent Bricks. It has three pieces, all built on Lake Base. >> [music] >> Managed sessions record each conversation turn by turn. Managed memory keeps what’s worth knowing about each user, preferences, facts, >> [music] >> and workflows. Memory pipelines connect the two. In the background, [music] they distill sessions into memory, merge duplicates,

00:33 and keep memory files easy to access. The Databricks research [music] team found that agents with memory passed three times as many tasks, using 31% fewer tokens. [music] Hooking it up is easy. Bind a session store and a memory store, then create a memory pipeline between the stores. >> [music] [snorts] >> Now your memories are stored in Lake Base, and Databricks automatically synthesizes new ones in the background. >> [music]

01:04 >> Your agent’s next conversation starts with what it learned from the last one. Agent [music] memory, now in beta on Agent Bricks. >> [music]