Graphon Labs Raises $35M Series A to Build Memory for AI Agents

Why existing approaches fall short

Graphon says RAG, GraphRAG and large context windows can retrieve information, but compress, fragment or lose context, relationships, history and timing as data grows and changes. Embedding-based search can also blur differences between near-identical phrases with opposite meanings, such as "refundable" and "non-refundable," or "70V" and "770V."

The company's approach rests on graphons, a mathematical framework for analyzing very large graphs, which its founder studied during doctoral research at the University of Pennsylvania's GRASP Lab. Graphon builds one large graph from all of a customer's data. A query first finds a few matches, and the graphon then guides retrieval toward connected information. The system ingests video, images, audio and documents and builds an ontology automatically, so it can search for "a short-sleeved blue shirt with five buttons and two pockets" rather than just "blue shirt," and show where a result came from.

Three products and benchmarks

Graphon Enterprise gives company data of any format governed memory and can be deployed hosted, in a Graphon-managed VPC, in a customer-managed VPC or on-premises. It holds SOC 2 Type II certification. Graphon Developer offers S3-style storage and retrieval for keyword, semantic and image search, plus agent queries across multiple documents. Graphon On-Device Memory runs private, persistent multimodal intelligence on phones, laptops, cameras and wearables.

On three long-context benchmarks, the company reports the highest accuracy among the systems evaluated: 93.9% on LoCoMo with 58% fewer tokens than loading entire conversations into the model, 90.2% on ConvoMem with 95% fewer tokens, and 81.7% on Cambridge researchers' ATM-Bench, 5.0 points above the next entry on the public leaderboard.

According to the company, customers already include the San Francisco 49ers and Hitachi. The new funding will go toward research and infrastructure to extend Persistent Intelligence beyond enterprises to developers and devices.