Graphon Labs Raises $35M Series A to Scale Persistent AI Memory

Graphon Labs has raised a $35 million Series A, led by RRE Ventures, to scale persistent-memory infrastructure for AI agents across enterprise, developer and on-device products. The round, announced November 10, 2026, brings the San Francisco startup's total funding to $43.3 million. HP, Energy Impact Partners, Hitachi Ventures and Gaia VC joined the round alongside RRE.

The bet is straightforward: today's AI systems forget. Chief executive Arbaaz Khan said the new capital will move Graphon from early experimentation into scaled enterprise deployments, open its infrastructure to developers, accelerate research and support on-device memory. That's a commercialization agenda, not just a research one — and it puts Graphon in a race with much larger players trying to solve the same problem.

Why persistent memory matters for AI agents

RRE general partner Raju Rishi framed the investment in a single sentence that captures the company's pitch: "Models can reason, and agents can act, but neither gets better if they cannot build off of prior knowledge."

That's the gap Graphon wants to fill. As AI agents move from chatbots to systems that take actions inside companies, they need to remember what happened last week, last quarter and last year. Current approaches don't handle that well, according to the company.

Graphon argues that retrieval-augmented generation, GraphRAG, context windows and conventional retrieval all compress or fragment relationships, history and temporal understanding as information grows and changes. In plain terms: the bigger your data gets, the more the existing tools lose the thread. Rishi's firm sees Graphon becoming a source of truth beneath agents — the memory layer everything else sits on.

How the Graphon architecture actually works

Graphon calls its approach Persistent Intelligence: AI that can continuously access, understand, remember and build on earlier information. The mechanics start with ingestion. The platform takes in video, images, audio and documents, then automatically builds an ontology — a structured representation of the data and its connections — while retaining links back to the original sources.

Khan's technical explanation splits the architecture into two pieces:

• A detailed graph representing information and relationships.
• A compressed "graphon" representation that guides retrieval through the larger structure.

Graphons are a mathematical framework for very large graphs, and they give the company both its name and its intellectual core. Khan, chief technology officer Clark Zhang and chief operating officer Deepak Mishra developed the underlying research during their PhDs at the University of Pennsylvania's GRASP Lab. The engineering goals are concrete: incorporate new information as it arrives, trace evidence back to its origin, and run at practical cost and speed.

The developer-facing side is deliberately unglamorous. Developers work with S3-style storage and search concepts — buckets, objects, files and queries — without needing to understand the graph mathematics underneath. The interface supports keyword search, semantic search, image search and agent-assisted questions across multiple sources. A quick lookup and a deep investigation draw on the same platform, so teams don't have to assemble a separate retrieval stack for each kind of question.

The takeaway: the hard math lives below the waterline, and the surface looks like familiar storage tooling.

What the benchmark numbers say

Graphon reports 81.7% accuracy on ATM-Bench, a University of Cambridge benchmark for question answering across video, audio, images and text. The company says that result sits 5.0 percentage points ahead of the next system on the public leaderboard.

It's a single benchmark, and it's self-reported. But multi-modal question answering — pulling an answer from a video, a document and an audio file at once — is exactly the task where fragmented retrieval breaks down. A five-point lead on a public leaderboard is the kind of number that gets enterprise architects to schedule a demo.

Three products, three different buyers

The architecture supports three products aimed at distinct markets:

• Graphon Enterprise — governed memory across enterprise data, hosted by Graphon, deployed in a Graphon-managed or customer-managed virtual private cloud, or installed on-premises.
• Graphon Developer — for builders working across data types without recreating the context stack themselves.
• Graphon On-Device Memory — for phones, laptops, cameras, wearables and other edge devices.

The deployment options matter for enterprise sales. Companies with strict data rules can keep everything on-premises, while others can use Graphon's cloud. HP's participation in the round suggests hardware and on-device use is a real part of the thesis, not an afterthought.

Who is actually using it?

The announcement names the San Francisco 49ers, Hitachi and other large enterprises as deployment partners, but the customer quotes reveal a more exploratory picture.

Costa Kladianos, a technology executive with the 49ers, said the organization is exploring applications such as uncovering sponsorship opportunities and supporting internal teams. That's a pilot, not a rollout. Hitachi Rail's Amaury Jourdan described an evaluation showing promise for contextual-knowledge discovery, automated data-source connections and improved AI inference in the company's industrial setting. Promising, but still an evaluation.

The earlier customer evidence goes back to the company's seed round. Graphon emerged from stealth on May 14, 2026 with an $8.3 million round led by Arvind Gupta of Novera Ventures. That announcement identified South Korean conglomerate GS as an early enterprise customer, with a representative citing applications including analysis of customer movement in convenience stores and CCTV-based construction-safety work.

The pattern across all three customers is early-stage: exploration, evaluation, pilot. Graphon's company profile lists July 2025 as its founding date, making this a company barely a year and a half old with a $43.3 million pile of capital to deploy.

What the $35 million has to prove

The Series A funds three separate tests at once. Graphon must turn exploratory enterprise work into scaled deployments. It must persuade outside developers to build on its memory layer rather than roll their own. And it must adapt the same architecture for private, resource-constrained on-device use — a very different engineering problem from a cloud deployment.

Each test has real competition. Every major AI lab and cloud vendor is working on memory for agents, and developers have a growing menu of open-source retrieval frameworks. Graphon's differentiator is the graphon mathematics and the claimed benchmark lead, plus deployment flexibility that runs from Graphon-hosted to fully on-premises.

Whether that's enough depends on execution over the next year. The company has the funding, the named customers and the benchmark numbers; now it has to show enterprise pilots converting into contracts and developers adopting a platform whose core innovation most of them will never see.