Databricks' MemCurrent Triples AI Agent Success While Using 12x Fewer Tokens
Databricks unveils MemCurrent, a persistent memory system that triples agent pass rates while cutting tokens 31% and latency 26% in production.
- Databricks introduces MemCurrent, a persistent memory architecture for production agents.
- Triples agent pass rate on internal benchmark while using 31% fewer tokens and running 26% faster.
- Combines episodic memory (raw traces) with derived memory (distilled files) via plugins, tools, and offline dreaming.
- Matches full-context accuracy on public benchmarks at roughly 12x fewer input tokens per question.
- Shipping in Agent Bricks Memory on Lakebase, with CLI quickstart under 10 minutes.
- Already powers memory in Databricks Genie natural-language data interface.
Databricks has introduced MemCurrent, a persistent memory architecture already powering Agent Bricks Memory and the Genie natural-language data interface. The system carries facts, preferences, corrections, and workflows across sessions. On Databricks’ 114-case enterprise benchmark, it more than tripled the pass rate of a no-memory agent while using 31% fewer tokens and completing tasks 26% faster.
Agents that begin every session without retained state must repeatedly discover the same information. Supplying the complete conversation history can prevent that relearning, but long prompts increase cost and latency while exposing the model to stale, conflicting, or irrelevant details. MemCurrent stores raw history alongside distilled knowledge, then retrieves only the material needed for the current task.
Two memory layers divide the work
Episodic memory preserves raw session traces. Derived memory extracts reusable knowledge from those traces and organizes it as files in a structured knowledge base. An automatically generated memory.md index describes the file structure so the agent can locate relevant material. Both layers reside in Lakebase, Databricks’ low-latency database service.
Explicit requests flow through the add_memory, update_memory, and delete_memory tools. A user instruction such as “Never write scripts in the /tmp directory” can therefore take effect during the current session. An asynchronous process called dreaming handles implicit learning after the session, including splitting oversized files, merging duplicates, and resolving conflicts.
At the start of each session, a plugin injects the memory index into the prompt. During execution, the agent can call search_memory to find relevant entries and fetch_memory to retrieve their contents.
| Pattern | Role | Advantage | Constraint |
|---|---|---|---|
| Session plugin | Loads the memory index at startup | Makes available knowledge visible immediately | Adds tokens to every session |
| Agent tools | Search, fetch, add, update, or delete memory | Retrieves specific information on demand | Depends on the agent choosing the right tool |
| Offline pipeline | Consolidates implicit learning in batches | Moves extraction work outside the live request | Introduces a delay before updates appear |
A benchmark drawn from workspace traffic
Many public memory benchmarks measure whether a model can recall facts from synthetic conversations. Databricks built an additional benchmark around enterprise actions such as interpreting table semantics, applying metric definitions, and following established workflows.
The company began with more than 12,000 Genie sessions and identified 3,000 candidates containing potentially useful memory signals. It then generated follow-up tasks and scoring rubrics. An adversarial judge checked each candidate against seven criteria, including whether the required knowledge was present, user-specific, necessary, natural, and verifiable. Manual review reduced the final set to 114 test cases spanning 29 users and 2,233 historical sessions.
- Operational tasks: Retrieved knowledge must guide multi-step actions, including data analysis and tool use.
- Varied memory conditions: Five memory types are crossed with 11 scenarios.
- Conflicting information: Cases include changed facts, outdated instructions, and knowledge accumulated across sessions.
- Control cases: Memory-independent questions test whether retrieval introduces unnecessary information.
The reported pass rate represents the share of tasks that satisfied their scoring rubrics, rather than simple fact-recall accuracy.
Similar accuracy with one-twelfth the context
Databricks compared MemCurrent with Mem0, PropMem, and OpenClaw on LoCoMo, LongMemEval, and ConvoMem. Each memory backend used the same underlying language model. The evaluation also included a no-memory floor and a full-context baseline that supplied the complete history.
| Configuration | Average score | Input tokens per question |
|---|---|---|
| MemCurrent | 88.1% | 8,400 |
| Full context | 87.8% | 103,000 |
MemCurrent recorded the highest score among the tested memory backends and used about 12 times fewer input tokens than full context. It also outscored the full-context baseline on the longest-history datasets, where additional prompt material can distract the agent from relevant evidence.
Results on the enterprise benchmark produced a larger separation between systems. Gold injection supplied the task’s required memory directly and served as an upper-bound reference.
| Configuration | Pass rate |
|---|---|
| No memory | 22.3% |
| Mem0 | 43.9% |
| MemCurrent | 68.4% |
| Gold injection | 78.6% |
MemCurrent closed more than 80% of the gap between no memory and gold injection. It also recorded the lowest latency and token use among the compared memory systems, at 252 seconds and 418,000 tokens. The evaluation used Claude Opus 4.6 for the agent and GPT 5.6 sol for memory processing.
Ablations isolate the useful pieces
Databricks separately disabled parts of the architecture to measure their contribution. Derived memory supplied most of the improvement, while episodic traces recovered details lost during distillation.
| Memory configuration | Pass rate |
|---|---|
| No memory | 22.3% |
| Episodic only | 35.8% |
| Derived only | 66.1% |
| Episodic and derived | 68.4% |
The agent searched episodic history in roughly one-third of test cases and passed 60% of the cases where it performed that search. Derived memory worked more consistently because the startup index exposed available knowledge before the task began.
A separate extraction ablation compared online tools with offline dreaming. Online tools reached a 53.9% pass rate, offline dreaming reached 63.5%, and combining them reached 66.1%. The two mechanisms cover different timing requirements: tools apply explicit instructions during a live session, while dreaming consolidates implicit patterns in the background.
Saved corrections shorten later runs
Two anonymized Genie sessions illustrate how retained knowledge changes execution. In one case, the agent generated correct SQL but failed to read the saved results file, then invented a table in its response. After the correction was stored, the agent handled the same failure pattern correctly in a later session.
Another session required 11 tool calls, eight subagents, and almost six minutes to answer a first-time request for the “top 5 SDP customers.” The agent saved the discovered workflow. A related follow-up later required two tool calls, one subagent, and 41 seconds.
Agent Bricks provides the shipping path
MemCurrent is available through Agent Bricks Memory, backed by Lakebase. The offering includes:
- a session store for episodic traces;
- a structured store for derived knowledge;
- the automatic dreaming pipeline;
- tracing and evaluation hooks for measuring memory’s effect; and
- a command-line tool that creates a sample agent in under 10 minutes.
Genie users receive the same memory architecture through the managed interface.
Deployment still requires workload-specific tests
The 114-case enterprise benchmark reflects one vendor’s products, users, and task distribution, while the public benchmark figures are also Databricks-reported. Teams evaluating the system should reproduce quality, latency, and token measurements with their own models, tools, prompts, and history lengths.
Production reviews should also cover user and tenant isolation, retention periods, deletion propagation, permissions for shared memory, recovery from incorrect entries, and monitoring for stale or conflicting knowledge. These controls become more consequential when a memory store persists across model upgrades or serves several agents.
MemCurrent treats memory as a service separate from any single agent or language model. That separation can preserve learned workflows during model changes and support controlled knowledge sharing across an organization, provided the surrounding access and lifecycle policies are equally durable.
Comments
