AI Agents Keep Forgetting, So Graphon Is Betting $35M on a Graph

AI Agents can write code, query databases, and perform multi-step tasks. But they often cannot remember what was said in the previous turn, let alone handle years of accumulated emails, images, videos, and documents inside a company. Graphon Labs wants to solve this, and it has just closed a $35 million Series A.

The company does not chunk data and then run vector retrieval. According to founder Arbaaz Khan in Business Insider, common AI search is like tearing a book into pages, crumpling them into a ball, and throwing them into a trash can. When you dig through the can, you only find fragments; the relationships between pages are gone. Graphon builds a graph spanning documents, emails, images, and videos, preserving the connections, provenance, and changes over time between pieces of information, without relying on a manual schema. When new data comes in, there is no need to redesign the ontology.

Arbaaz Khan and his co-founders previously worked on robotics memory at NASA JPL, Google DeepMind, Meta, and Nvidia. That background explains why they have kept focusing on 'Why does AI still forget?'. This round was led by RRE Ventures, with participation from Energy Impact Partners, Hitachi Ventures, Gaia VC, and HP. The company has 20 people, and its customers include the San Francisco 49ers and HP. Six months ago, it had just raised about $8 million in seed funding.

Based on the figures the company has released, benchmarks are its strongest marketing material. On LoCoMo, Graphon achieves 94.0% accuracy using 7.9K tokens. By comparison, stuffing the full history into the model costs 19K tokens and yields 90.4%. On ATM-Bench, Graphon V1 ranks first at 81.71%. In DolphinBench, it completes 504 of 600 tasks, with 84.0% accuracy and a total cost of $11.84. The best result for the same model on the public leaderboard is 70.7%, at a cost of $96.21. The company also reports up to 96.2% on LongMemEval and 73.0 on BEAM-10M.

The 94% stands out, but the 7.9K tokens are more worth watching. If a memory system relies only on stuffing more history into context, cost and latency will rise with it. Graphon is talking about a different curve: using fewer tokens to achieve higher accuracy. It packages indexing, search, Q&A, and research into an API; its storage layer is S3-compatible; and the company says you can point Claude, GPT, or another off-the-shelf agent at it, or directly ingest 100K PDFs without changing the ontology.

But enterprise memory is not just about accuracy. Someone on X asked how permissions would be handled when a single graph spans company emails, with the agent only able to retrieve content the questioner could already see. That question is harder than benchmark scores. Another concern is that schema-less graphs claim new data will not break the structure, but real-world data changes, conflicts, and expires. Another user asked how already-built memories stay fresh after documents are modified. Graphon's ATM-Bench Hard score is 65.6, and it is still awaiting leaderboard review after submission. Benchmarks themselves also need time to be validated.

Graphon's narrative wins on the cost-accuracy frontier, not on simply topping leaderboards. It pulls the memory problem back from 'stuff in a little more context' to 'how data connects, how it changes, and who can see it.' If this graph can truly stay updated and enforce permission isolation in enterprise data, agents will go from interns who restart every time to colleagues who remember the ins and outs of a project. If not, the graph is just another data silo—except this time $35 million is riding on it.

Reference links:

• Original X post: https://x.com/omarsar0/status/2120204152479400099
• Business Insider: https://www.businessinsider.com/graphon-pitch-deck-funding-ai-agents-amnesia-2026-10
• ATM-Bench leaderboard: https://atmbench.github.io/leaderboard.html
• Graphon benchmarks page: https://www.graphonlabs.ai/#benchmarks