A San Francisco startup is betting that autonomous AI agents need something most companies haven't given them yet: a working memory that actually understands how organizations operate.
Zeroset announced Sunday it raised $5.2 million in pre-seed funding co-led by Gradient Ventures and 2048 Ventures, with Leblon Capital joining the round. The company declined to share its valuation. Its product, Nebula, represents an attempt to solve what investors and engineers increasingly describe as the reliability gap in enterprise AI automation.
The pitch centers on a technical distinction that may seem subtle but carries weight in implementation. Nebula isn't designed to retrieve information when an AI agent asks a question. Instead, it models the flow of work across an organization's systems, creating what Zeroset calls a "state layer" that helps agents behave consistently over time rather than simply answer queries.
"Models off the shelf have no understanding of the dynamics of how an enterprise actually works," CEO Akshat Kannan said.
That lack of context, Kannan argues, is why many early enterprise AI deployments stall after handling initial tasks. An agent might successfully answer customer emails but struggle when the workflow requires coordinating across Slack, GitHub, and a CRM system. Nebula aims to map those handoffs, turning observed patterns into what the company describes as repeatable playbooks.
The technology connects to Microsoft 365, SharePoint, Outlook, Teams, GitHub, and workplace messaging platforms, according to coverage by Business Insider. Under the hood, Zeroset built what it describes as a custom hierarchical vector graph that traces how work moves through an organization. The company plans hybrid pricing combining license fees with usage-based charges tied to data volume and agent activity.
Young Team, Bold Claims
Kannan left Stanford to launch Zeroset, his LinkedIn profile shows. Co-founder William Zhang departed UT Austin, where he'd previously interned at Amazon working on retrieval-augmented generation systems. Before joining forces with Kannan, Zhang founded Fynopsis, another retrieval-focused venture. The company lists 2–10 employees on LinkedIn.
Performance benchmarks provided by Gradient Ventures in its investment announcement suggest Nebula delivers 20% higher retrieval accuracy than competitors including Mem0, Supermemory, and basic RAG implementations on LongMemEval, a standard evaluation. According to the same investor blog post, median retrieval clocks in under 50 milliseconds with fewer tokens per query. On BPI Challenge datasets, which test process mining capabilities, a Nebula-based system outperformed recorded implementations on next-activity prediction by more than 10%, according to Gradient's claims—though these benchmarks have not been independently verified.
The company offers REST API access with TypeScript and Python SDKs. Documentation indicates the API handles memory storage and search with citation tracking, plus support for Model Context Protocol integration.
Crowded Territory, Different Angle
The enterprise AI memory space has attracted multiple approaches. Mem0 provides both open-source and managed memory layers for AI agents, integrating across agent frameworks. Zep operates a managed temporal knowledge graph with enterprise governance features. Both focus primarily on retrieval.
Zeroset draws a line between its approach and competitors' in a research essay the company published recently. Retrieval answers questions, the argument goes, but state determines behavior. That distinction matters when an agent needs to decide not just what information to surface but what action to take next in a multi-step workflow.
"Until it is captured somewhere an agent can use, enterprise agents will keep stalling a few steps in," Gradient Ventures wrote Sunday.
2048 Ventures highlighted another ambition in its investment thesis: training customer-specific models on structured traces of workflow patterns. The firm calls these "enterprise world models" that learn which actions actually move business processes forward, though practical deployment of such models remains in early stages.
Nebula is currently in closed research preview. Zeroset is hiring researchers and systems engineers, according to Sunday's announcement.
The funding comes as venture investors pour capital into infrastructure for agentic AI, betting that autonomous systems will need specialized tooling to work reliably in complex organizational environments. Whether memory and state management emerges as a distinct product category or gets absorbed into broader agent platforms remains an open question. Zeroset is wagering it's the former.
