Agent Memory entered Beta on Agent Bricks, built on the MemCurrent architecture and Lakebase

The main development in October 2026 was Agent Memory, which Databricks released in Beta on Agent Bricks on 20 October [3]. It gives agents a Lakebase-backed store for facts, preferences, and workflows, and it is built on the MemCurrent architecture from Databricks Research [3]. A launch video the day before described session stores and memory stores connected by automated memory pipelines, so that agents keep context across sessions and use fewer tokens [4]. In Databricks' internal benchmarks on enterprise tasks, Agent Memory tripled the agent pass rate [3].

A separate technical write-up described MemCurrent as combining agent-written memory with an offline "dreaming" pipeline. That pipeline consolidates, deduplicates, and resolves conflicts in a structured knowledge base [2]. The same post said the architecture also powers Genie, and it reported results on an internal benchmark built from real Genie conversations [2]. These benchmark figures came from Databricks itself, and the month's material did not include independent verification.

Other material placed memory inside a wider platform. A 12 October overview presented Agent Bricks as a developer platform with flexible hosting, a unified API for model routing, secure sandboxes, managed memory, and tools for real-time tracing and evaluation [6]. A demo on 19 October walked through a six step workflow covering vector search indexing, agent as code, MLflow tracing, Unity Catalog registration, evaluation, and deployment with review apps [5]. On 1 October, a demo showed a governed store operations agent using custom Model Context Protocol servers, MLflow traces, security policies, and budget controls through Unity AI Gateway [8]. A community post also covered running production agents on enterprise data [1].

On the community side, one discussion asked for recommendations on limiting spend on a pay-as-you-go account [7].