Memory Scaling for AI Agents

Databricks Research introduced MemAlign, a memory framework for AI agents that stores past interactions as episodic memories and uses an LLM to distill them into generalized semantic rules, which are then retrieved at inference time to guide the agent. Tested on Databricks Genie Spaces—a natural-language-to-SQL interface—the goal was to measure how agent performance scales as more memory (from both curated labeled examples and raw user conversation logs) is fed to the agent. As memory grew, the average reasoning steps per example dropped from ~20 to ~5, with the agent learning to retrieve relevant context directly rather than exploring the database from scratch. A key finding was that uncurated user interactions, filtered only by an automated LLM judge, can substitute for costly hand-engineered domain instructions, pointing toward agents that improve continuously from normal usage and scale beyond the limitations of human annotation.

Related

• Powering clinical research innovation: How TriNetX uses Databricks to accelerate drug development
• Database Branching in Postgres: Git-Style Workflows with Databricks Lakebase
• The Next Era of the Open Lakehouse: Apache Iceberg™ v3 in Public Preview on Databricks
• How Zalando built a unified data foundation for AI and analytics on Databricks

Source: industry