Graphon Labs Raises $35M Series A for AI Agent Memory

Company knowledge has a change log, whether the AI system keeps it or not. Graphon Labs is building persistent memory that connects information with its sources, relationships and history, putting the maintenance of that context at the center of its AI infrastructure business.

The San Francisco company announced a $35M Series A on November 10, 2026, led by RRE Ventures with HP, Energy Impact Partners, Hitachi Ventures and Gaia VC participating. The investment supports broader enterprise deployments, developer access and research toward on-device memory, extending the audience for a technology first introduced to enterprise buyers.

Graphon reports $43.3M in total funding, including its earlier $8.3M seed. The commercial question running through that expansion is who takes responsibility for keeping an agent's knowledge useful as the organization, its records and the devices accessing them change.

What Graphon Labs' Series A finances

Graphon's financing brings more resources to an architectural bet already under development. The company wants a persistent representation of information to serve agents across enterprise environments, developer applications and personal devices. Each setting changes the practical conditions under which the same underlying memory has to operate, including where the data can live and how it reaches the system asking a question.

The Series A announcement describes Graphon Enterprise, Graphon Developer and Graphon On-Device Memory as three product directions. Enterprise options include hosted infrastructure, Graphon-managed or customer-managed virtual private clouds, and on-premises deployments. Expansion to devices is part of the company's stated ambition; the financing announcement supplies no universal delivery date for that roadmap.

The capital history is comparatively straightforward. Graphon emerged from stealth with $8.3M in seed funding in May 2026, in a round led by Arvind Gupta of Novera Ventures. The new $35M is the Series A amount, while $43.3M is the company-reported cumulative total. The primary announcements use dollar signs without specifying a denomination, and a valuation was not disclosed.

That funding sequence gives Graphon a wider deployment agenda within months of its public debut. It also gives the founders several audiences to serve: enterprise teams governing shared information, developers building applications around it, and eventually individuals whose data lives close to the device producing it.

The founders' bet on connected information

Graphon's current team page identifies Arbaaz Khan as founder and CEO, Clark Zhang as co-founder and CTO, and Deepak Mishra as co-founder and COO. The company reports a July 2025 founding and traces the underlying research to work at Penn's GRASP Lab. Its technology uses graphons, mathematical representations of large graphs, to organize relationships across multimodal information.

Arbaaz Khan's account of the company's motivation begins with a particularly demanding environment for plausible answers: Amazon customer service, where he worked as an applied scientist. A policy can change while a model's training remains fixed. Retrieving a relevant passage helps, but the exception or warning that gives that passage meaning may live somewhere else.

That history makes Graphon's positioning easier to understand. The company is selling a way to keep connections available before an agent starts forming its answer. In its description, video, images, audio and documents contribute to a persistent relational memory rather than remaining separate collections of searchable material.

For a buyer, the consequential detail is the relationship between a statement and the circumstances around it. An answer assembled from company information needs to retain enough of that relationship to remain usable when the business changes. Graphon's commercial opportunity sits inside the engineering work required to make that continuity dependable.

Enterprise experience and the limits of comparison

Graphon says its technology is deployed with the San Francisco 49ers, Hitachi and other enterprises. The customer statements give those relationships useful shape: the 49ers discuss exploring sponsorship opportunities, while Hitachi Rail describes evaluating connections across complex industrial information. Those accounts establish different areas of interest without supplying revenue, contract size or a common measure of customer outcomes.

The earlier seed announcement offers another application through GS, the South Korean conglomerate. A GS executive described work involving customer movement in convenience stores and CCTV analysis at construction sites. These examples help explain why Graphon emphasizes several data types: the information needed for a business question can extend beyond a written document.

Graphon's announcement also reports results on LoCoMo, ConvoMem and ATM-Bench. Its live site now displays updated figures, so the versions deserve to remain distinct. Benchmark comparisons are useful evidence about a specified evaluation; translating them into a general promise about a customer's data would require more than a leaderboard position.

The same care applies to the language around AGI in the company's announcement. Persistent context is Graphon's chosen research direction toward that ambition. Enterprise buyers can evaluate the nearer job in front of them: whether the system retrieves the relevant evidence, preserves its relationships and fits the controls around their own information.

A shared memory layer gains more customers to answer to

Developer access changes the scope of the infrastructure sale. A team building an agent may otherwise assemble storage, retrieval, source citations and repeated evaluation itself. Graphon's current platform exposes storage, search, answer and research APIs, offering a route to buying more of that continuing work from one provider.

That creates a relationship extending beyond the first successful query. New documents arrive, instructions change and applications ask questions their builders did not anticipate during setup. A memory provider becomes part of the customer's process for keeping information usable, while the customer retains responsibility for deciding what the resulting answer is allowed to support.

On-device expansion adds another set of constraints to that relationship. Information may need to stay near a machine, camera or individual user, with a different compute budget from a hosted enterprise environment. Graphon's plan to span those settings gives the Series A a specific purpose: carrying its representation of connected information into places where rebuilding an enterprise retrieval stack would be a considerable ask.

For Arbaaz Khan, Clark Zhang and Deepak Mishra, the financing widens the range of people who can depend on the memory they are building. A developer's next application and a device's next recording will add more history to that system, along with another reason to keep the connections intact.