Graphon Labs raises $35M to give AI agents memory that keeps its context

Agents can write code and carry out multistep tasks, but they still lose track of past conversations and struggle with large, changing troves of company data[3]. Graphon says common approaches such as retrieval-augmented generation lose context, relationships and timing as data grows[2].

"The old way, using vector search, is like searching through the pages of a torn-up book. We take those pages and turn them into a connected map, unlocking seamless memory search," founder and CEO Arbaaz Khan said[1].

The system is built on graphons, a mathematical framework for very large graphs that Khan and co-founder Clark Zhang used for robot memory during their PhDs[2]. It preserves what an agent knows, what it has done, what changed and what happened as a result, with sources and timing intact[1]. Off-the-shelf models such as Claude or GPT can be pointed at it directly[2].

Graphon reports 93.9% on LoCoMo, a test of memory over long conversations, using 58% fewer tokens than loading the whole conversation, and 81.7% on the University of Cambridge ATM-Bench, 5 points ahead of the next system on the public leaderboard[2]. Those figures are company-reported.

The San Francisco 49ers, Hitachi and other large enterprises already use the software[1]. The 20-person company will spend mainly on engineers and researchers, plus some sales staff, and stay focused on enterprise before pushing further into devices[2]. It sells three products: Graphon Enterprise, Graphon Developer and Graphon On-Device Memory[1].

As agents run for hours or days rather than single chats, memory becomes the limit on how much they improve. The large labs are adding memory to their own assistants, so Graphon's bet is that a model-agnostic memory layer, owned by the company rather than the model vendor, is worth paying for.