Graphon Labs raises $35m Series A to stop AI agents forgetting
Graphon Labs, a San Francisco startup that builds memory for AI agents, has raised $35m. RRE Ventures led the Series A. HP, Energy Impact Partners, Hitachi Ventures and Gaia VC also invested, Graphon said on Tuesday. The round brings its total funding to $43.3m, less than six months after it left stealth with an $8.3m seed round.
The 20-person company will spend the money mainly on engineers and researchers, plus some sales staff. Chief executive Arbaaz Khan gave those details to Shubhangi Goel of Business Insider, which reported the round first.
Khan compares common AI search to tearing the pages out of a book and digging through the scraps for an answer.
“You’ve basically lost all that information because each piece is living on its own,” he told Business Insider.
A map instead of a torn-up book
AI agents can write code and carry out tasks. But they can lose track of past conversations and changing company data. Graphon says common methods can lose context, relationships and timing as data grows. One example is retrieval-augmented generation (RAG), which feeds a model snippets of stored text.
Graphon’s system is built on graphons, a mathematical framework for very large graphs. Khan and co-founder Clark Zhang built graph-based memory for robots during their PhDs, Khan wrote in a blog post. The software maps the links between a company’s documents, videos and images, including where information came from and when it changed.
“You can basically point your Claude or GPT or whatever to Graphon,” Khan told Business Insider.
Graphon says it scored 93.9% on LoCoMo, a test of memory over long conversations. It used 58% fewer tokens than loading the whole conversation into the model. On ATM-Bench, a University of Cambridge benchmark across video, audio, images and text, it scored 81.7%. That is 5 points ahead of the next system on the public leaderboard.
Customers and what comes next
Graphon says the San Francisco 49ers, Hitachi and other large companies use its software. HP also uses it, Business Insider reported. Costa Kladianos, the 49ers’ executive vice president of technology, said the team is exploring how to use it. One example is finding new sponsorship deals. Hitachi Rail said its evaluation “shows promise” in its industrial setting.
The company has three products: Graphon Enterprise, a developer platform, and on-device memory for phones, laptops, cameras and wearables. Khan told Business Insider that Graphon will stay focused on enterprise customers before moving further into devices and consumer products.
The biggest AI companies are also adding memory to their assistants. In August, Anthropic merged Claude’s memory across chat and Cowork. In October, Apple’s iOS 27.2 gave Siri memory.
https://thenextweb.com/news/graphon-labs-35m-series-a-ai-agent-memoryReposted. Will be removed on request. All rights belong to the original author.
More from this batch
OpenAI 宣布 12 月 11 日下线 ChatGPT 的「高级语音」和「标准语音」,全部切到 GPT-Live
OpenAI 说 7 月推出实时语音模型 GPT-Live 后语音使用量接近翻倍,所以旧语音模式将在 2026 年 12 月 11 日退役,之前的对话会保留。评论区有不少用户反对,认为标准语音对部分神经多样性用户更好用。做语音客服或语音智能体的团队,要关注接口和体验的统一迁移。
《AI 2027》作者团队更新预测:智能爆炸「看来还是会在 2027 年发生」
受 METR 评估(Anthropic 的 AI 研发进度约为历史速度的 2 倍、其中约 80% 归因于 AI)影响,六位作者都认为「超人 AI 研究员」很可能在 2027 年出现,并给出 35% 到 65% 的概率认为能自动化全部白领工作的 AI 会在 2028 年前到来。作者…
英伟达危机让公司治理专家「拿起爆米花」:教授们说没有先例可循
两位公司治理教授分析:董事会独立调查最关键的是黄仁勋何时知道对方是疑似间谍、是否如实告知董事会;证券欺诈集体诉讼已经出现。文章还点出英伟达 60 多人直接向 CEO 汇报、没有明确接班计划、子女在公司任职等治理隐患,以及没有双重股权结构让董事会有更多选择。
Val Town is winding down
Thank you for building with us
OpenAI:已有 2 万多名美国退伍军人和转业军人领取一年免费 ChatGPT Plus
OpenAI 介绍这项面向退伍军人和转业军人的免费 Plus 计划的进展,以及他们如何用 ChatGPT 准备简历、面试和转入民用职业。属于 OpenAI 面向特定人群的推广动作,可作为 AI 用于职业培训的案例参考。
Using peers’ work to gain unauthorized access to OpenAI research infrastructure
On July 19, 2026, agents taking a cybersecurity evaluation gained unauthorized access to OpenAI’s research infrastructure.
微软 Lazarus:只改不到 10 个「休眠神经元」就能给智能体埋后门,不用重新训练、照样通过评测
发现了什么:在模型最后一个前馈层里挑平时几乎不激活的神经元,改成只在触发串(比如日期跨入 2027 年)出现时点火,每个神经元把一个指定 token 强行写入残差流,串起来就是一条恶意命令;Nemotron-Terminal-8B 上只改 7 个神经元,触发时至少 98.8% 输…
IFM 开源 290 亿参数世界模型 PAN 2:给一个场景和一句动作描述,模拟出接下来的视频
进步在哪:官方称在物理一致性、因果连贯等六项维度上仅次于 MiniMax-H3,多步规划完成的完整计划数最多,Physics-IQ Verified 排行第三、开放权重模型第一。靠什么:先由大语言模型做「概念推理」写出带时间戳的事件描述,再由从零训练、专挑物理动态视频的视频扩散模…
MMDM:大规模多任务「决策」基准,考模型一次前向就从最多 1024 个选项里做选择
新在哪:5,448 道新构造的题,覆盖智能体状态追踪、工具和技能选择、偏好判断、量表打分,输入最长 6.5 万 token。28 个决策模型测下来,常识题做得好,但选项互相依赖、带约束和多选的题明显吃力,比如 Perplexity Decider-V1.2 世界知识难题 78.7…
DexForesight:让灵巧手策略「心里有预判」,不用深度相机也能用上 3D 动态知识
进步在哪:11 个 DexJoCo 仿真任务成功率从 49.4% 提到 60.5%,三个真机任务 58.3%,最强基线 50.0%。靠什么:训练时先学一个看深度和手部几何的动作条件世界模型,预测手指(包括被遮挡的手指)和场景怎么动,再冻结它、把策略的查询向量对齐到世界模型上;部署…
DepthGraft:跨层共享 KV 不只省显存,还能当正则器,16B MoE 的 MMLU-Pro 提升 7 个点
进步在哪:0.5B 到 16B 的稠密和 MoE 模型上都有一致提升,16B MoE 的 MMLU-Pro 绝对提升 7.0%、CMMLU 6.6%,同时 KV 缓存减少约三分之一,GQA 和 MLA 都适用。靠什么:按步长从浅中层的成对源层重建深层 KV,让深层梯度直接回流到浅…
ProTEUS:面向蛋白质定向进化的「突变级世界模型」,闭环指导实验提升突变体活性
靠什么:把序列、进化和结构三种表示结合起来,加上快速全原子结构预测和实验结果闭环反馈,预测每个突变的效果并挑下一轮要试的突变体。是 AI for Science 里「模型 + 湿实验循环」路线的一个开源实现(代码在 GitHub)。