Menu
Crypto sections
The San Francisco startup's Series A, led by RRE Ventures, bets that enterprise AI needs connected memory, not bigger context windows
Tue 10 Nov 2026 · 12:18 EST · 3 min read
In AI
Share
AI agents are getting smarter, but many still have the memory of a goldfish at a crowded party. Graphon AI just raised $35 million to fix that.
The San Francisco startup closed a Series A round on November 10, 2026, to build what it calls persistent relational memory for AI agents. The goal is to let those agents reason across huge piles of enterprise data without forgetting what they saw a few documents ago.
The details of the raise
RRE Ventures led the round. HP and Hitachi Ventures also participated, which gives the cap table a noticeably corporate flavor.
The timing is aggressive. Graphon secured $8.3 million in seed funding on May 14, 2026, which means the Series A arrived roughly six months later.
The company itself is young. Graphon was founded in July 2025, so it went from founding to a $35 million Series A in a little over a year.
Novera Ventures was among the seed investors, alongside several other tech-focused backers. The Series A counts as a follow-on to that earlier round, so the momentum carried over from one raise to the next.
On the customer side, Graphon is actively working with GS Group in South Korea as one of its early deployments. The company also has emerging partnerships involving HP and the San Francisco 49ers, two names that rarely appear in the same sentence.
What Graphon actually builds
Graphon describes its platform as a “pre-model intelligence layer,” meaning it sits underneath the AI model and prepares information before the model ever touches it.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
We respect your privacy. Unsubscribe anytime.
The platform pulls in many kinds of data at once. Documents, videos, images, and logs all go into the system.
From that mix, it builds a knowledge graph that is self-supervised and schema-free. Put more simply, the software figures out how pieces of information connect on its own, without a human first drawing up a rigid structure for it to follow.
The underlying math leans on graphon functions, the mathematical concept that gives the company its name. Those functions help model relationships across very large networks of connected data.
The result, according to the company, is that agents running on models like GPT or Claude can tap into integrated, time-aware information. Instead of grabbing isolated scraps, they see how facts relate and how they change over time.
Why context windows aren’t enough
Most AI models work within a context window, the amount of text they can hold in view at once. Once information falls outside that window, the model effectively stops seeing it.
A common workaround is retrieval-augmented generation, usually shortened to RAG. In that setup, the system fetches relevant chunks of data and hands them to the model when it needs them.
The catch is that those chunks often arrive disconnected from each other. Graphon is positioning its knowledge graph as the answer to that problem. By storing relationships rather than fragments, it aims to let agents reason over long stretches of multimodal enterprise data.
The company points to benchmark results to back up the pitch. Graphon ranked first on ATM-Bench, scored 94% on LoCoMo, and hit 84% on DolphinBench.
What this means for enterprise AI
Graphon’s model-agnostic approach is a notable part of its pitch. If its memory layer works with models like GPT or Claude, enterprises aren’t forced to bet on a single AI provider to use it.
The investor lineup also hints at the go-to-market strategy. Having HP and Hitachi Ventures in the round, plus an early deployment with GS Group, suggests Graphon is aiming squarely at big corporate customers rather than hobbyists.
The range of early partners is worth watching as well. A Korean conglomerate, a hardware maker, and an NFL franchise represent very different data problems, so they could serve as a test of how versatile the platform really is.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.
The news about what you hold
Follow your stocks and crypto. Read the news about them from hundreds of publishers in Pro.
Privacy policy
