Databricks details MemCurrent, the memory design behind Agent Bricks Memory

Databricks published MemCurrent, an agent memory architecture pairing raw session logs with distilled memory files consolidated by an offline 'dreaming' pass.

Databricks has published MemCurrent (October 20, 2026), the memory architecture behind its new Agent Bricks Memory and Genie. If you're hand-rolling memory for an agent, it gives a concrete answer to what to store, when to write it and how to keep it fresh, and that answer is useful whether or not you run Databricks.

The design has two layers. Episodic memory is an append-only log of raw session traces. Derived memory holds distilled learnings as files, with an auto-generated memory.md index that a plugin injects at session start. The agent gets add_memory, update_memory and delete_memory tools for explicit requests, and search_memory and fetch_memory for recall. An offline "dreaming" pipeline extracts implicit learnings after each session, splits oversized files, merges duplicates and resolves conflicts. The post doesn't detail how conflicts are resolved.

On Databricks' internal benchmark of 114 cases built from Genie sessions, the pass rate went from 22.3% with no memory to 68.4%, with 31% fewer tokens and 26% faster runs. Mem0 scored 43.9% on the same test. In Databricks' ablations, offline dreaming alone (63.5%) beat online tools alone (53.9%). ZIPMEX notes that these are Databricks' own figures, and that the public-benchmark and internal runs used different models, so the two sets of results don't compare directly.

Agent Memory is in Beta on Lakebase, Databricks' serverless Postgres. The announcement doesn't mention pricing or say whether you can use it outside Databricks. If you need memory that runs on your own infrastructure, Particula Tech describes Mem0 as Apache-2.0 and self-hostable. The two-layer, consolidate-offline pattern carries over to any stack.

Sources