engram
The challenges are growing as companies deploy AI tools and autonomous agents into more business functions. Technology often lacks a lasting understanding of the organizations it serves. Despite advances in language models at scale, most AI systems repeatedly process the same documents, relearn the same context, and rebuild organizational knowledge with each new interaction, creating both cost and efficiency concerns for businesses operating at scale.
san francisco-based on engram aims to address this problem with a new approach to enterprise AI memory. The company came out of stealth today. $98 million with funding from General Catalyst, Kleiner Perkins, Sequoia Capital, Factory, Modern, Amplify Partners, NeoWith support from a group of technology leaders including With Co-founder and CEO Asaf Rappaport, OpenAI co-founder Andrei Karpathyand AI researcher Peter Abbeer.
Rather than relying solely on a search system to expand its context window, Engram trains its models to proactively examine an organization’s information, creating what it calls a compact and continuously improving memory layer for each customer. According to the company, this approach allows the model to retain organizational knowledge over time, while reducing the computational resources required to answer questions and complete tasks.
“The AI that people are using today is mostly improvising what they know about their organizations in real time,” he said. Dan BidermanCEO and co-founder of Engram. “If you can predict interactions in advance, you can build memories in advance rather than recreating the context for each conversation.”
The startup’s technology is rooted in academic research across machine learning, neuroscience, and AI memory systems. Biderman completed postdoctoral research at Stanford University under machine learning researcher Chris Le, who is also a co-founder of Engram. Before that, he completed his Ph.D. Graduated from Columbia University Center for Theoretical Neuroscience.
The founding team includes researchers from Stanford, Berkeley, and Cornell with expertise in AI learning, search systems, and model memory. Sabri Eyuboglu, a PhD and chief technology officer at Stanford University, has developed Cartridge, a way to convert large collections of documents into reusable AI memory. Co-founder Jesse Lin holds a PhD from Berkeley. The former Meta FAIR researcher developed active reading, a technique designed to help models learn information more deeply. Jack Morris, a Cornell University PhD FAIR alumnus, is known for his work on retrieval and memory in large-scale language models, while Stanford University professor Scott Linderman specializes in state-space models, an emerging alternative architecture for processing long sequences of information.
According to Engram, this technology can significantly reduce the number of tokens needed to operate enterprise AI systems. Eyuboglu said traditional AI models can generate huge internal memory representations when processing long documents, which can lead to higher infrastructure costs and slower performance.
“We train the model in such a way that it learns once in advance and compresses everything it learns into a compact memory that can be reused for every query,” Eyuboglu said.
The company has already entered the market with several high-profile partnerships. Engram is working with Microsoft to evaluate its technology within the Microsoft 365 environment, with a focus on improving efficiency and enabling AI systems to better understand an organization’s context. This collaboration includes access to GPU infrastructure through Dapple and Microsoft Azure, providing resources to train and scale its models.
“Our customers have amassed extraordinary knowledge within Microsoft 365, and we are just beginning to take advantage of what it can do for them,” he said. Jason Greifcorporate vice president of AI Partner Catalyst at Microsoft. He said Engram’s approach could help organizations create and maintain AI memory while supporting the development of long-running enterprise agents.
The startup has also partnered with workplace software provider Notion and legal AI company Harvey. The companies are investigating how learned memory can improve the performance of persistent AI agents operating across large and complex knowledge environments.
Simon Last, co-founder of Notion, said the company has been testing Engram’s models within its custom AI agents and is seeing promising efficiency gains because the system already understands the underlying workspace, rather than rebuilding the context for each interaction. Gabe Pereyra, co-founder and president of Harvey, said the technology has the potential to help organizations transform large amounts of unstructured information into long-term organizational knowledge for AI-driven workflows.
Investors view the memory layer as a potentially critical component of enterprise AI infrastructure. Hemant Taneja, CEO of General Catalyst, said organizations in sectors such as healthcare, legal services and financial services are amassing large amounts of proprietary knowledge while facing rising costs of operating AI.
For Engram, the larger goal is to give businesses greater ownership of the intelligence created through the deployment of AI. The company argues that organizational knowledge should become a unique asset that improves over time, rather than being a value that accrues primarily to the model provider.
Founded in San Francisco, Engram says its platform allows organizations to build AI systems that become more specialized as they are used, creating a durable organizational memory designed specifically for each company.
