Graphon Labs Raises $35 Million to Give AI Agents Persistent Memory, Claiming 93.9% on LoCoMo With 58% Fewer Tokens

Graphon Labs, an applied AI lab building memory infrastructure for AI agents, announced Tuesday a $35 million Series A led by RRE Ventures, with participation from HP, Energy Impact Partners, Hitachi Ventures and Gaia VC. The round comes less than six months after the company emerged from stealth with an $8.3 million seed, bringing total funding to $43.3 million.

The company's pitch targets a weakness every agent deployment eventually hits: models can reason and agents can act, but neither gets better if it cannot build on what happened before. Retrieval-augmented generation, GraphRAG and ever-larger context windows can pull information back in, but Graphon argues they compress, fragment or lose relationships, history and temporal context as data grows.

Graphon's answer comes from mathematics rather than a bigger index. Its architecture is built on graphons — a framework for analyzing very large graphs that the founders developed during their PhDs — and it reworks the whole stack from storage through search, retrieval, reasoning and memory. CEO Arbaaz Khan likens traditional vector search to "searching through the pages of a torn-up book," while Graphon turns those pages into a connected map that preserves what an agent knew, did, and what changed as a result.

The company reports the highest accuracy among evaluated systems on three long-context benchmarks, all company-reported figures: 93.9% on LoCoMo while using 58% fewer tokens than loading the full conversation into the model, 90.2% on ConvoMem with 95% fewer tokens, and 81.7% on ATM-Bench, a Cambridge-run benchmark spanning video, audio, images and text, where it says it leads the next system on the public leaderboard by 5.0 percentage points. The LoCoMo and ConvoMem comparisons use baseline results published by Databricks, and none of the numbers have been independently verified.

Deployment is already real, if narrow. Graphon says its technology is in production with the San Francisco 49ers and Hitachi, and the press release names Hitachi Rail among evaluators exploring context discovery and automated data-source connections. The product line spans Graphon Enterprise for governed corporate memory, Graphon Developer for builders, and Graphon On-Device Memory for phones, laptops and wearables. The founding team comes from Amazon, Meta, Google, Apple and NVIDIA.

The funding thesis is simple: if agents are going to run continuously rather than answer one question at a time, memory becomes the layer that everything else sits on. "We see it becoming the source of truth beneath every agent," said Raju Rishi, the RRE general partner who led the round. Whether Graphon's graph-first approach beats cheaper vector-plus-context stacks at scale is the open question, but investors are now paying $35 million to find out.