Agent memory

Managed, Lakebase-backed state for agents built on any framework, part of Agent Bricks. Agent sessions keep the conversation history of one interaction, agent memory keeps facts, preferences and workflows that an agent recalls in later conversations through combined full-text and semantic search, and memory pipelines use an LLM to distill sessions into memories on demand or about every 24 hours. Stores are workspace-scoped with store-level access control, and Databricks says the design follows its MemCurrent memory research, which also powers memory in Genie. The first version, launched in Beta on 2026-06-23 as Unity Catalog memory stores, is now legacy and deprecated in the docs. Lakebase-backed memory and sessions joined on 2026-09-16 and memory pipelines on 2026-10-12. This row was previously listed as Managed agent memory.

Status and dates are extracted from public sources only. We don't see the private AE-shared roadmap. If something looks wrong, tell us.