Not everyone agrees on what an “AI agent” actually is, but they are all still the rage. At a broad level, these so-called “agents” promise to go several steps beyond a mere chatbot, making decisions and taking actions on people’s behalf. Some might help you do your online shopping; others might make factories more efficient — at varying degrees of autonomy.
It’s against such a backdrop that fledgling AI startup, Across AI, is coming out of stealth to develop a “dynamic memory system” for complex enterprise workflows. And it’s spearheaded by a founder who relatively recently sold his previous startup to IBM.
Across AI targets the likes of chief revenue officers and sales teams with a platform that connects with all their internal and external enterprise data sources. It then creates a shared “agentic memory” that can be used to identify and qualify fresh sales opportunities, spot risks, and suggest questions sales teams should be asking their customers.
“Sales teams often struggle with obtaining and utilizing the right information when they need it — whether that’s information about products, customers, competitors or optimal processes,” Across AI’s co-founder and CEO, Steven Mih (pictured above, center), told TechCrunch. “Critical knowledge often gets bottlenecked among a few experts or buried in vast amounts of unstructured data, leading to inefficiencies, delayed decisions, and costly errors. Existing AI solutions often fail to address this issue because they lack deep integration and contextual understanding, treating all data equally without the ability to prioritize or adapt to new information.”
Mih was previously co-founder and CEO of Ahana, a Google Ventures-backed company that built commercial services atop Presto, the open-source SQL query engine that spun out of Facebook in 2013. Mih sold Ahana to IBM last year for an undisclosed amount, and after a 14-month stint at the tech giant, Mih jumped ship in July to begin work on his latest startup.
He joined up with Dr. Niloufar Salehi (pictured above, left) and Dr. Afshin Nikzad (pictured above, right), renowned professors from UC Berkeley and Stanford University, respectively, who have carried out research on ways to improve the efficacy of AI systems in “high-stakes” settings.
Across AI is still embryonic — it’s refining its product with design partners in private. As it works toward a commercial launch in 2025, the company has now raised $5.75 million in a seed round of funding co-led by Bobby Yazdani‘s Cota Capital, and Village Global, a venture capital firm that counts Bill Gates, Mark Zuckerberg, Jeff Bezos, and Reid Hoffman among its backers.
Building memories
Across AI will be a web app and chatbot that connects to various parts of the enterprise stack — CRM systems, communication and collaboration tools, calendars, and all the rest — to build its memory and develop contextual understanding. It will then be on-hand to help wherever a user is working.
“By showing up where users already are, for example in Slack or [Microsoft] Team’s app, Across AI apps don’t break the user’s flow and instead provide just-in-time assistance in the context of the user’s existing workflows,” Mih said.
This memory, the company says, “continuously adapts,” and only retains what it deems to be relevant information while discarding outdated data. This raises questions around how it could determine what’s relevant, as this is highly dependent on the context and requirements of the people who will use it.
Mih says it achieves this by developing a “deep understanding of the workflow context.”
“The system actively tracks, timestamps, and monitors information updates, recognizing when data becomes outdated or conflicts with new information,” he said. “Unlike traditional AI systems that treat all data equally, our agentic memory system prioritizes information based on contextual importance. Where possible, the apps keep the inferences up to date themselves. Where ambiguity exists, determinations are escalated to a relevant person, such as a sales manager or product manager.”
Of course, enterprises have been slow to adopt generative AI, as data privacy and security are still core concerns. The last thing a company wants to do is funnel all its proprietary and sensitive data off to a third party, which then does God knows what with it. As such, Mih says that data security is a “foundational aspect” of the startup’s agentic memory platform.
“Our memory system operates within the company’s secure environment, maintains access control over sensitive information, and does not expose data to external models for training,” Mih said. “We plan to offer both SaaS and cloud-premises deployment options to meet enterprise security and compliance requirements.”
There are subtle synergies between Mih’s previous startup and his latest venture. Ahana was all about enabling users to query vast amounts of data via Presto, with Ahana taking care of all the complexities around infrastructure setup and maintenance. Across AI is addressing the same problem, but through a different lens.
“I believe that a core differentiator for AI application companies will be their ability to help users analyze large amounts of data, quickly — that’s exactly what we specialized in at Ahana,” Mih said. “This experience deepened my understanding of the challenges enterprises face in making sense of complex data ecosystems that are often siloed and hard to navigate.”
