Zeroset
Category: AI Infrastructure
Applied AI research lab building Nebula, a memory and state layer that gives enterprise AI agents a live, traceable model of how a company's workflows actually run. Zeroset was founded in 2026. The company is led by Akshat Kannan. Based in San Francisco, United States. Team size: 1-10. Total funding raised: $5.2M. Latest round: Pre-Seed. Key investors include Gradient Ventures, 2048 Ventures, Leblon Capital.
- Founded
- 2026
- Headquarters
- San Francisco, United States
- Team size
- 1-10
- Total funding
- $5.2M
Value proposition
Nebula connects to the systems a company already uses (Microsoft 365, SharePoint, Outlook, Teams, GitHub, messaging tools) and turns tickets, messages, approvals, handoffs and tool calls into typed, queryable workflow traces. Agents get the smallest current state they need for each action, so they complete more work, repeat fewer mistakes and use less context.
Products and solutions
Nebula, a state layer for AI agents, currently in closed research preview. It includes Nebula Workflows, which traces handoffs across people and agents and turns them into repeatable playbooks. Published research includes the Atlas v0 agent-memory benchmark and the 'Memory Is Not Retrieval' memo. Upcoming research includes the JASPER world model for workflow traces, the SOLAR architecture and λBench.
Unique value
Nebula stores state, not just vectors. It resolves entities, time, contradictions, ownership and provenance into workflow traces, instead of returning top-k similar chunks the way vector DBs and memory providers do.
Target customer
Enterprises and AI-native companies running long-lived, multi-step AI agents, initially in manufacturing, supply-chain operations and financial research.
Industries served
Manufacturing, Supply chain operations, Financial research, Enterprise software / AI-native companies
Technology advantage
Uses a custom hierarchical vector graph architecture. According to investor 2048 Ventures, Nebula beat Mem0's retrieval accuracy on LongMemEval by about 20%, with sub-50ms p50 retrieval at lower token cost. The roadmap is customer-specific 'enterprise world models' trained on structured traces (JEPA-inspired). Enterprise features include single-tenant deployment, customer-managed keys and SIEM audit logs; a SOC 2 Type II audit is in progress.
How they differentiate
Positions itself as a workflow state layer rather than memory retrieval. It tracks how entities, facts and processes change over time, with sources, and returns minimal context. It also aims to train enterprise-specific world models on top of that data, while competitors focus on chunk retrieval.
Main competitors
Mem0, Zep
Key partnerships
Infrastructure on AWS. Model providers are Azure OpenAI, OpenAI, Google Cloud Vertex AI, Groq and Modal, per the Zeroset security page.
Major milestones
Founded January 2026. Published the Atlas v0 agent-memory benchmark in January 2026. Emerged from stealth with a $5.2M pre-seed round on 2026-10-06. SOC 2 Type II audit in progress.
Growth metrics
5 employees at funding announcement (Oct 2026). Product is in closed research preview. One peak training run cost about $100K over two weeks. Hiring for 4+ San Francisco roles (research, distributed systems, Rust systems, product).
Market positioning
Early-stage agent infrastructure startup in the AI agent memory/state category. It targets enterprise deployments with licence plus usage-based pricing and subsidizes the first workflow proof for qualified teams.
Geographic focus
United States / global enterprise
About Akshat Kannan
Stanford University CS (dropped out in 2026 to work on Zeroset full time); co-founded Zeroset at age 17-18. Confirmed as current co-founder & CEO by Business Insider (2026-10-06) and Economic Times (2026-10-10).
Latest news about Zeroset
More AI Infrastructure companies
Official website: https://zeroset.com