Six months ago Graphon came out of stealth with an $8.3 million seed and a claim that AI’s context problem should be solved outside the model. On November 10 the San Francisco company, now called Graphon Labs, announced a $35 million Series A led by RRE Ventures. HP, Energy Impact Partners, Hitachi Ventures and Gaia VC joined. Total funding is $43.3 million.
Memory outside the model
Large language models forget. Everything they know about your data has to fit in a context window, and retrieval systems like RAG pull back isolated snippets without the relationships between them. Graphon’s approach is to process the data first. Its platform ingests video, images, audio, documents and logs, discovers how the pieces connect and stores that as persistent relational memory any model or agent can query.
The name comes from the maths. A graphon is a continuous representation of a graph, a framework co-invented by the company’s UC Berkeley advisors. Founder and CEO Arbaaz Khan, CTO Clark Zhang and COO Deepak Mishra came from labs at Amazon, Meta, Google, Nvidia and others.
The company reports 93.9% accuracy on the LoCoMo long-conversation memory benchmark and 90.2% on ConvoMem, with token use cut by 58% and 95% respectively in its comparisons. Fewer tokens means lower inference bills, which is the practical argument for buyers.
Where it’s being used
Graphon names enterprise content management, industrial monitoring, agent workflows and on-device AI on phones, cameras and wearables. Customers quoted include the San Francisco 49ers and Hitachi Rail. Deployment options cover Graphon’s cloud, a customer’s private cloud and on-device installs, which matters for regulated industries that won’t send records to a third party.
The bigger idea is that you don’t get to better AI only by training bigger models. You also get there by giving existing models a better picture of the world they’re reasoning about. Graphon calls that persistent intelligence. Investors are now paying four times the seed for the bet.