New feature
Beta
Managed agent memory
Managed agent memory is a Databricks ai / ml capability, introduced June 2026.
A fully managed, Lakebase-backed store that gives agents durable long-term memory across conversations, usable from agents built on any framework.
- For all the agent talk, recall is old-school full-text BM25 search, returning a relevance-ranked top-N of at most 100 entries.
- A memory store can never forget its own name - display_name is immutable after creation, and only the description can be changed.
- It moved house within a quarter - the June 2026 Beta note kept memories in Unity Catalog memory stores, while the September note describes it as a Lakebase-backed store.
Limitations: Beta. Long-term memory only (short-term history is managed agent sessions); search is full-text BM25 returning up to 100 entries; access control is per store, with no per-entry or per-actor control; a store's display_name cannot be changed after creation. During the preview you pay for the underlying Lakebase instance.
Open in REbricked →- Category
- AI / ML
- Also known as
- agent memory, long-term agent memory
- Verified
- 2026-10-06