Zeroset Raises $5.2M for Enterprise Agent Memory

Zeroset raised a $5.2M pre-seed round to build Nebula, a state layer designed to help enterprise AI agents understand how work changes across systems. Gradient and 2048 Ventures co-led the financing, with Leblon Capital participating in the October 6, 2026 announcement.

Co-founders Akshat Kannan, CEO, and William Zhang are tackling a problem that gets harder as agents move beyond isolated tasks. The model may know how to draft, search, classify, and reason, but the company-specific answer to “what should happen next?” often depends on a changing chain of approvals, ownership, exceptions, handoffs, and unwritten precedent.

Nebula remains in closed research preview, so the funding is backing an infrastructure thesis before broad product availability. Zeroset plans to use the capital for model training, early enterprise deployments, and hiring researchers and systems engineers while it tests whether workflow state can become a durable layer in the enterprise-agent stack.

What Zeroset Is Building

Enterprise software records activity in pieces. A ticket may contain the request, Slack may contain an approval, a CRM may reflect the commercial consequence, and an employee may remember why the exception was granted. An agent that retrieves only the most similar documents can still miss the sequence and current state connecting those records.

Nebula is designed to capture tickets, messages, approvals, handoffs, tool calls, and decision records, then organize them into a live model of the business. According to Zeroset and its investors, the system tracks what happened, what changed, what remains unresolved, and how one person's work relates to another's. A long-running agent can query that state instead of reconstructing the company from raw history on every run.

Zeroset describes the approach as an enterprise world model, borrowing a useful distinction from robotics. A capable policy still needs a representation of the environment where it acts. In enterprise automation, that environment includes private workflows and changing rules that no foundation model learned from the public web.

Why the State Layer Matters

The enterprise-agent market has concentrated enormous investment on model capability and tool access. Those advances make agents better at discrete work, but they do not automatically tell an agent whether last week's approval still applies, which policy version governed a decision, or whether an unresolved exception should stop the next action.

That gap is partly technical and partly organizational. State has to preserve entities, time, ownership, provenance, contradictions, and the reason behind a decision. Zeroset argues that this operating knowledge should remain controlled by the enterprise, allowing companies to use frontier models or train open models without giving up the accumulated record of how their work gets done.

This positioning puts Nebula upstream of task execution. The model can choose or perform an action, but the state layer supplies the compact, current context that makes the action relevant to this company and this moment. That is a harder product claim than document search because customers will judge it through completed workflows, intervention rates, repeated errors, latency, and cost.

The $5.2M Pre-Seed and Its Economics

The financing was announced by Zeroset as a pre-seed and described by 2048 Ventures as an inception round. Gradient and 2048 Ventures co-led the $5.2M investment, while Leblon Capital participated. No valuation, prior institutional financing, or additional participating investor was publicly verified.

Business Insider reported that Zeroset had five employees at announcement and planned to expand access progressively rather than launch Nebula broadly. The company intends to combine a license component with usage pricing tied to data volume and agent activity. It is initially focused on enterprises and AI-native companies running long-lived agents in areas including manufacturing, supply-chain operations, and financial research.

The training budget makes the size of the round more legible. Kannan told Business Insider that one peak model-training run cost about $100K over two weeks. For an early company still refining product behavior and enterprise deployment, the capital has to cover research ambition, compute bills, customer integration, and the systems work required to earn trust inside consequential workflows.

Founders, Product Evidence, and What Is Still Unproven

Kannan and Zhang bring an unusually early founder story to an infrastructure problem. Business Insider reported that Kannan left Stanford and Zhang left the University of Texas to work on Zeroset; 2048 Ventures also credits Zhang with experience building production retrieval-augmented generation pipelines at Amazon and founding Fynopsis. The verified public record supports Kannan as co-founder and CEO and Zhang as co-founder, without a confirmed CTO title.

Zeroset has published technical work arguing that memory should be maintained as evolving state rather than treated only as retrieval. Its Atlas benchmark tests world modeling, temporal reasoning, procedural knowledge, preference learning, and other capabilities that the company believes long-running agents require. Gradient and 2048 Ventures report that Nebula achieved 20% higher LongMemEval retrieval accuracy than comparison systems with sub-50ms median retrieval and lower token use.

Those results are useful evidence of the company's technical direction, but they are produced or repeated by Zeroset and its investors. They are not an independent audit of customer outcomes, production reliability, or the ability to complete complex enterprise workflows without human supervision. Nebula's closed-preview status keeps that distinction important.

What the Round Signals for Enterprise AI

Zeroset is entering a busy infrastructure market that includes memory providers such as Mem0 and Zep, along with orchestration, context, integration, observability, and security vendors trying to make agents reliable in production. DevCuration has also tracked funding for durable agent infrastructure and long-running AI agents, evidence that investors are breaking the enterprise-agent problem into distinct layers.

Zeroset's bet is that state becomes one of those layers. The company is not selling a more articulate model; it is building the machinery that tells a model what changed inside a particular business and why the next action belongs in the workflow. If that layer works, model choice can remain flexible while company-specific operating knowledge compounds.

The next phase will be decided inside early deployments, where elegant memory architecture meets messy permissions, contradictory records, sensitive data, and employees who have been carrying the real workflow in their heads. The $5.2M round gives Zeroset time to move that knowledge into software, but the company's opportunity will grow only as Nebula proves it can keep enterprise agents aligned with work that refuses to stand still.