Success Triples AI Performance Across 114 Business Tasks

Databricks, a company that sells data and AI tools, said its MemCurrent software helped AI assistants raise success on 114 business tasks from 22.3% to 68.4% while finishing faster and using less text.

Key Takeaways

Databricks MemCurrent raised agent success from 22.3% to 68.4% on 114 enterprise tasks
Agents finished tasks 26% faster while using 31% fewer tokens, Databricks said
The internal benchmark reported scores. No memory 22.3%, episodic 35.8%, derived 66.1%, and both together 68.4%
Databricks said MemCurrent matched full prior-conversation performance on three public benchmarks while using 12 times fewer tokens

The company posted the results on X and linked to a longer writeup that said agents also finished tasks 26% faster with 31% fewer tokens. Databricks sells data and AI infrastructure that large companies use to build analytics and machine learning systems.

A second article linked from the post laid out the internal benchmark.

It broke performance into four cases, no memory at 22.3%, episodic memory at 35.8%, derived memory at 66.1%, and both together at 68.4%.

Databricks MemCurrent Attacks Agent Forgetfulness

The argument is simple. Most agents can reason inside one session for success, but they keep forgetting user preferences, past actions and internal company patterns once the session ends.

Databricks MemCurrent packages that missing layer as two forms of memory.

Episodic memory stores raw interactions. Derived memory stores patterns, preferences and workflows for success inferred from those interactions.

That split matters because raw history alone gets expensive fast.

Passing whole chat logs back into a model burns tokens, slows execution and raises the odds that useful evidence gets buried under irrelevant detail.

The writeup said MemCurrent matched the effect of giving an agent all prior conversations on three public benchmarks, but did so with 12 times fewer tokens. It also said the system beat three open memory tools, OpenClaw, Mem0 and PropMem, on both accuracy and efficiency.

Also Read: Software Firm Gives Companies 1 Rulebook For AI Across Microsoft And Amazon

From Prompt Windows To Persistent Workflows

The timing fits a broader shift in enterprise AI.

Through most of 2025 and early 2026, companies focused on bigger context windows, hoping larger prompts would solve recall.

That only goes so far. A context window is the amount of text a model can read in one pass.

It is not a working memory that persists across weeks of sales calls, support tickets or compliance reviews.

The new pitch from Databricks is that enterprise agents need continuity more than raw size. If that view holds, the spending race may tilt from larger base models toward middleware that decides what to remember, what to compress and what to retrieve.

That also creates a cleaner commercial wedge.

Companies can swap foundation models later, but a memory layer tied to workflows and internal knowledge may prove harder to replace once deployed.

Where Databricks MemCurrent Could Face Limits

Databricks did not release the full benchmark methodology in the X post, so outside replication is still the next test. Internal enterprise tasks can be useful, but buyers will want to see performance across real deployment logs.

Another open question is failure mode.

Derived memory sounds powerful, but any system that summarizes behavior can also preserve wrong assumptions. A bad preference, once stored, may be reused over and over.

The company is also entering a crowded area.

A wave of startups and open-source projects are trying to become the memory layer for agents, especially in support, coding and knowledge work.

How Databricks Splits Agent Memory For Success

Databricks MemCurrent matters because it separates memory into two jobs that enterprise teams already understand. Episodic memory keeps the raw record.

Derived memory turns that record into reusable facts.

For a support team using Databricks, episodic memory would keep the exact ticket history with a customer. Derived memory would store that the customer always wants weekend maintenance windows and escalates database issues to one internal lead.

The agent does not need every old exchange for success each time. It needs the right facts.

That changes the economics.

Passing full chat history into a model makes token bills rise line by line. A retrieval layer picks only the snippets or summaries the model needs for the next step.

Databricks said that approach matched full-history performance with 12 times fewer tokens, which is the kind of number procurement teams notice.

What to watch next is outside validation. If developers reproduce the 68.4% result or beat the 12x token savings on public workloads, Databricks MemCurrent stops being a benchmark claim and starts becoming a buying category.