Akshat Kannan dropped out of Stanford at 18. William Zhang left the University of Texas at 19. In October 2026, their startup, Zeroset, secured $5.2 million to build a memory for machines. The bet is that the next bottleneck for AI agents in the enterprise is not raw intelligence, but a persistent record of what happened last time [Business Insider, October 2026].
Their product, Nebula, is a closed research preview positioned as a memory and state layer. It connects to enterprise systems like Microsoft 365, GitHub, and Teams. Its job is not to index static documents, but to track how information, decisions, and activity change over time [Business Insider, October 2026]. The intended result is a temporal workflow history. An agent can reference it to repeat a successful procurement procedure or avoid a past mistake in a supply chain alert. The wedge is sequence.
The temporal workflow wedge
Most enterprise AI tooling today focuses on retrieval. It finds a document, a policy, or a piece of code. Zeroset's founders argue that the harder problem is context. What were the steps that led to a decision? Who was involved in the Slack thread that approved a vendor change? Which version of a financial model was used before the quarterly report was finalized?
Nebula aims to capture that sequence. It logs activity across connected tools, building a timeline of events that an autonomous agent can query. The initial target customers are enterprises and AI-native companies running long-horizon agents, with a focus on manufacturing, supply-chain operations, and financial research [Business Insider, October 2026]. For a logistics agent rerouting shipments, the memory of a past port delay and the alternative chosen could be the difference between a suggestion and an executable plan.
The team and the round
The founding story is a study in early conviction. Kannan, the CEO, was 18 when he left Stanford in early 2026 to start the company. Zhang, his co-founder, had dropped out of the University of Texas the previous year [Business Insider, October 2026]. By the time of their pre-seed announcement, the team counted five employees [Business Insider, October 2026]. Their current job postings seek researchers and distributed systems engineers fluent in Rust, signaling a build toward infrastructure-grade reliability [LinkedIn job posting].
The $5.2 million pre-seed round closed in October 2026. It was co-led by Gradient, Google's AI-focused venture fund, and 2048 Ventures, with participation from Leblon Capital [Business Insider, October 2026]. The investor syndicate suggests a belief in the foundational layer thesis. Gradient's check, in particular, is a vote on the technical ambition behind a memory system for agents.
The company said it would use the capital to hire researchers and engineers, support early enterprise deployments, and pay for model training [Business Insider, October 2026]. The planned commercial model combines a license with usage-based pricing tied to data volume and agent activity.
| Round | Amount | Lead Investors | Date |
|---|---|---|---|
| Pre-seed | $5.2M | Gradient, 2048 Ventures | Oct 2026 |
Where the wheels could come off
The ambition is large, and the path is narrow. Zeroset is betting that enterprises will grant a deep, persistent integration to a startup to log sensitive workflow data. It is also betting that the value of a temporal memory will be clear and billable before larger platforms bake similar capabilities into their own agent ecosystems.
The risks are not hypothetical. They are the standard pressures on any infrastructure startup trying to carve out a new layer.
- Integration depth. The value proposition collapses without deep, reliable connections to core enterprise systems like ERP, CRM, and communication tools. Each new connector is an engineering lift and a sales hurdle.
- The platform question. Microsoft, Salesforce, and other incumbents with vast workflow data could decide to build or buy their own agent memory layers, leveraging their inherent access.
- Proving the ROI. The company's stated pricing model ties cost to data volume and agent activity [Business Insider, October 2026]. This requires customers to see a direct, measurable improvement in agent performance or human efficiency attributable to the memory layer,a complex attribution challenge.
Zeroset's answer, for now, is focus. By keeping Nebula in a closed research preview and targeting specific verticals like manufacturing and financial research, it can iterate with early design partners to prove the use case before a broad launch.
The next twelve months
The immediate roadmap is execution. The funding provides an 18- to 24-month runway for a team of this size to move from research preview to a generally available product. The key milestones will be technical,scaling the data layer,and commercial,landing the first handful of paid enterprise contracts.
The Gradient and 2048 Ventures-led $5.2 million pre-seed is a substantial opening act. It sets a high bar for what the company must prove: that autonomous agents need a dedicated memory layer, and that Zeroset can be the company that builds it. The question for 2027 is whether the first enterprise deployments will show agents making better decisions because they remember the last one.
Sources
- [Business Insider, October 2026] AI agents don't always understand how companies work. This startup raised $5.2 million to solve that. | https://www.businessinsider.com/zeroset-startup-funding-ai-agents-gradient-ventures-enterprise-2026-10
- [Zeroset LinkedIn] Zeroset LinkedIn | https://www.linkedin.com/company/zeroset
- [LinkedIn job posting] MTS · Engineering · Systems (Rust) at Zeroset | https://www.linkedin.com/jobs/view/mts-%C2%B7-engineering-%C2%B7-systems-rust-at-zeroset-4438470453