Loading post…
DatabricksSirui Sun, Nihit DesaiAnnouncement
Introducing Agent Memory on Databricks
Summary
Databricks announces Agent Memory on its Agent Bricks platform, now in Beta. Agent Memory improves agent quality by storing critical facts, preferences, and workflows that agents should remember, reducing both cost and latency. It is built on the MemCurrent architecture from Databricks Research, which combines context logging plugins, on-demand tool access, and a background dreaming pipeline to consolidate memories, avoiding the cost of replaying full histories or building bespoke memory systems, all served on Lakebase for low latency. In internal benchmarks, agents with memory tripled pass rates on real enterprise tasks while using 31% fewer tokens compared to stateless agents. Agent Memory joins other Agent Bricks capabilities including hosting, Sandbox, Unity Gateway, pre-built MCP servers, and tracing and evaluation, providing a complete platform for building production agents at scale.
Context
Stateless agents in production struggle with two problems: they rediscover facts and workflows each session, driving up costs and latency, and they forget user preferences and corrections, requiring repeated input. Developers often replay full history, leading to roughly 12x more input tokens per question, or build custom memory on object storage or vector databases, which introduces complexity in extraction, deduplication, and retrieval. Agent Memory on Databricks addresses these issues natively.
Approach / What changed
Agent Memory on Databricks uses three components: managed sessions that record conversation history in native format, managed memory that stores important facts and preferences retrieved via hybrid keyword and semantic search, and memory pipelines that distill sessions into long-term memory in the background, removing duplicates and updating contradictions. All data is stored on Lakebase, Databricks’ serverless Postgres, for low-latency access and scalability. This implements the MemCurrent architecture from Databricks Research.
Takeaways
Agent Memory is based on the MemCurrent architecture, using three access patterns: context-logging plugins, on-demand tool calls, and a background dreaming pipeline, all served on Lakebase.
Internal benchmarks show Agent Memory tripled agent pass rates on enterprise tasks while reducing token usage by 31%, compared to stateless agents.
Agent Memory integrates with the larger Agent Bricks platform, including agent hosting, Sandbox, Unity Gateway, pre-built MCP servers, tracing, and evaluation.
FAQ
What problems does Agent Memory solve?
It solves the problems of stateless agents repeatedly rediscovering facts and forgetting user preferences. By storing durable information, it reduces costs and latency while personalizing interactions across sessions.
What infrastructure does Agent Memory use?
Agent Memory runs on Lakebase, Databricks' serverless PostgreSQL, provisioned in the user's workspace. This ensures low latency, high reliability, and scalable storage for memories and sessions.
How can I add Agent Memory to an existing agent?
You can use the Agent Bricks CLI and follow the tutorial in the documentation to add memory and sessions to an existing agent. The quickstart guide also shows how to create a new agent with memory in minutes.
Source and freshness
- Source
- databricks.com
- Published
- Added to Yomu
Summaries are generated from the original by a language model. Check the original before relying on it.
Keep reading
Related reading
The Good Stuff
Paid picks we use ourselves. Links carry a referral code — a subscription through one earns us a commission at no extra cost to you.
The best engineering reads, once a week
One useful email. Unsubscribe anytime.