Databricks ships Agent Memory as agent memory hits production

Databricks ships Agent Memory

Databricks launched Agent Memory on Agent Bricks, its agent developer platform, and put it into beta on October 20, 2026. The pitch targets a concrete production pain: stateless agents re-derive what they already learned — which table holds the usage data, what an internal acronym means, which join actually works — in every single session. In the company's internal benchmarks, the memory layer triples agent pass rates on real enterprise tasks while using 31% fewer tokens. Those numbers are Databricks's own; treat them as the vendor's claim, not an independent result. m:…7684

The system is built on MemCurrent, the research architecture pioneered by Databricks Research, and runs on Lakebase, Databricks' serverless Postgres, which the company says gives it reliability and low latency for production agents. There is also an explicit control plane: developers can read, export, and delete the entries an agent keeps about each user, so the memory layer does not become an opaque dossier. m:…7684

MemCurrent: two kinds of memory

MemCurrent splits memory into two types: episodic memory, which stores raw session traces, and derived memory, which holds the distilled learnings from those traces. Three access patterns feed it — plugins that log and load context automatically, tools the agent calls on demand, and an offline dreaming pipeline that consolidates memory in the background while the agent is idle. m:…2809 m:…7684

On Databricks's own numbers, derived memory does the heavy lifting: derived memory alone lifts the pass rate to 66.1%, and MemCurrent performs on par with full-context replay at roughly 12x fewer input tokens per question, while running 26% faster than agents with no memory. The company says MemCurrent scores highest among memory backends on open-source benchmarks; its internal test puts Mem0 at a 43.9% pass rate with higher latency and cost. All of this rests on Databricks's own 114-test-case benchmark, so the comparison is vendor-run throughout. m:…2809

Memory needs a clock

Mem0 took a different angle the same day: time. Its bitemporal reasoning tracks two independent timelines — when the event actually happened in the real world, and when the system learned about it. The failure mode is concrete: import an old conversation without its original timestamps, and a memory like "last weekend" gets resolved against today's server clock, permanently wrong, because the original conversation date is gone after the import. m:…5320

When a fact gets replaced, Mem0 keeps the old one, marked as superseded, so agents can query both the current truth and the history behind it. A "latest_only" search flag retrieves just the present, memory decay re-ranks stale memories rather than deleting them, and expiration dates hide memories from search after their time is up. m:…5320

Cheaper memory, and the price of remembering

Meanwhile researchers are driving memory costs down. In a preprint, the SEEM policy for long-video agent memory — produced by an automated multi-objective code search framework — reports 32x lower average token use and 91x shorter average response time than prior state-of-the-art methods, reaching 65.7% average accuracy across the EgoLifeQA, Ego-R1, and Video-MME benchmarks. These results are not peer-reviewed. m:…1979

But the TokenWorth framework argues memory itself has a cost. The authors show that retrieving a relevant but ineffective memory can consume scarce context and hinder downstream execution, so TokenWorth learns query-dependent memory worth from sparse execution feedback, achieving the highest task success rate with the fewest output tokens across three LifelongAgentBench environments. In the paper's framing, effective lifelong adaptation depends not on remembering more, but on learning what is worth remembering. m:…8510