Graphon Labs raises $35M to give AI agents a memory that doesn't fall apart

Graphon Labs, a San Francisco startup building memory for AI agents, said on Nov. 10, 2026, that it raised a $35 million Series A led by RRE Ventures, bringing its total funding to about $43 million less than six months after its seed announcement.

Every agent demo looks smart until it forgets what you told it Tuesday. Memory is becoming the real battleground — Anthropic and Apple are both bolting it onto their assistants — and Graphon is betting a startup can do it better as a layer any model plugs into. The benchmarks are self-reported; the 49ers using it to hunt sponsors is the more honest test.

Graphon Labs, a San Francisco company building memory for AI agents, said it raised a $35 million Series A led by RRE Ventures. A Series A is an early private funding round, usually the first sizable one after a product is already in the world. The Next Web puts total funding at $43.3 million. The company says $43 million. Both figures land less than six months after Graphon left stealth with an $8.3 million seed. The announcement is dated Nov. 10, 2026. The round size and the lead are the company's. The $43.3 million total is The Next Web's.

HP, Energy Impact Partners, Hitachi Ventures, and Gaia VC also invested. Those names are The Next Web's account of the round.

The company has 20 people. Chief executive Arbaaz Khan told Business Insider the money goes mainly to engineers and researchers, plus some sales staff. The Next Web carried that.

The software is built on graphons, a mathematical framework for very large graphs. A graph, here, is a map of things and the links between them, rather than a pile of separate files. Khan and co-founder Clark Zhang built graph-based memory for robots during their PhDs. Graphon maps the links among a company's documents, videos, and images, including where a piece of information came from and when it changed. Khan argues that the usual method, chopping a document into short chunks and searching the chunks, drops the context around a line. His example is "can offer a refund" set next to "can't offer a refund." Embeddings are the number lists a model uses to compare text. Once both lines are embeddings, they look almost the same. Those lines are his, on the company blog.

His analogy for that search is tearing the pages out of a book and digging through the scraps. "You can basically point your Claude or GPT or whatever to Graphon," he told Business Insider. The quotation is his. The Next Web carried it.

Graphon published its own scores. On LoCoMo, a test of memory across a long conversation, it says 93.9 percent, using 58 percent fewer tokens than loading the full conversation. A token is a chunk of text the model counts. On ATM-Bench, a University of Cambridge test that covers more than text, it says 81.7 percent, 5 points ahead of the next system on the public leaderboard. Those numbers are the company's claims. An outside lab did not sign them.

The Next Web, citing Business Insider, names three users: the San Francisco 49ers, Hitachi, and HP. Costa Kladianos, executive vice president of technology for the 49ers, said the team is exploring uses such as finding sponsorship deals. Hitachi Rail said its evaluation "shows promise." HP appears on the investor list as well. Those customer lines are theirs, as reported.

There are three products. Graphon Enterprise is the one for companies. A developer platform is the one for people building on it. On-device memory is meant for phones, laptops, cameras, and wearables. The company is focusing on enterprise sales first, meaning other businesses, before a product on a personal phone. Those lines are the company's.

The co-founders are Arbaaz Khan, chief executive, previously an applied scientist on Amazon's customer service team; Clark Zhang, chief technology officer; and Deepak Mishra, chief operating officer. Those roles are the company's, as The Next Web reported them.

The picture is Deepak Mishra, Arbaaz Khan, and Clark Zhang sitting on a tan leather couch. The credit is Graphon Labs, as carried by The Next Web.

In plain terms, a 20-person San Francisco company raised $35 million, led by RRE Ventures, so an AI agent can remember a company's files as a map of links instead of a pile of scraps. The test scores are Graphon's. The 49ers looking for sponsors with it is a customer trying the tool, not a published score.