Engram Raises $98M to Build an AI Memory Layer for Enterprise Models

San Francisco-based Engram has emerged from stealth with $98M in funding to build what it calls a learned memory layer for enterprise AI. The company was co-founded by CEO Dan Biderman, CTO Sabri Eyuboglu, co-founders Chris Ré, Jessy Lin, Jack Morris, and Scott Linderman. Investors include General Catalyst, Kleiner Perkins, Sequoia Capital, Factory, Modern Capital, Amplify Partners, Neo, and SV Angel, with Assaf Rappaport, Andrej Karpathy, and Pieter Abbeel participating as investors and advisors.

The company's premise is simple but ambitious. Instead of forcing AI systems to repeatedly process the same organizational knowledge, Engram trains models to internalize that knowledge so they can operate with dramatically fewer tokens while maintaining performance. The funding reflects a broader shift in enterprise AI infrastructure, where investors are increasingly backing technologies that make AI cheaper, faster, and more practical to deploy rather than simply building larger foundation models.

What Happened

Memory is a funny thing. Humans forget birthdays, passwords, and where they left the coffee mug 5 minutes ago. AI forgets the 200-page policy manual you fed it yesterday, then politely asks you to upload it again while burning through tokens like a teenager discovering a parent's credit card. Efficient? Not exactly.

That is the problem Engram walked straight into, and investors didn't just nod politely. They committed $98M to the vision. CEO and co-founder Dan Biderman, CTO and co-founder Sabri Eyuboglu, and co-founders Chris Ré, Jessy Lin, Jack Morris, and Scott Linderman believe the next breakthrough in enterprise AI will come from memory, not just reasoning. General Catalyst, Kleiner Perkins, Sequoia Capital, Factory, Modern Capital, Amplify Partners, Neo, and SV Angel backed that conviction with capital, while Assaf Rappaport, Andrej Karpathy, and Pieter Abbeel reinforce the technical credibility surrounding the company.

The announcement positions Engram among a growing class of AI infrastructure startups focused on improving efficiency rather than simply increasing model size. As enterprise adoption accelerates, reducing inference costs and improving long-term knowledge retention are becoming strategic priorities instead of engineering niceties.

Why This Matters

The name Engram is doing some heavy lifting. In neuroscience, an engram is the physical trace of memory, making it an appropriate name for a company focused on helping AI retain knowledge instead of repeatedly relearning it. Most enterprise AI systems continuously process the same organizational context. Imagine hiring the smartest employee in the building, then forcing that person to reread the employee handbook before answering every question. That isn't intelligence. It's expensive repetition.

According to Engram, its learned memory layer trains models on an organization's context so they can operate using just 1% to 10% of the tokens, with certain workflows achieving up to a 100x reduction. If those efficiency gains hold in production environments, the implications extend beyond cloud costs. They reshape the economics of enterprise AI deployment by making persistent AI agents significantly more practical while allowing organizations to build AI systems that accumulate knowledge instead of repeatedly rebuilding context.

Market Context

Enterprise AI has entered a different phase. The early race centered on building larger large language models (LLMs). Today's competition increasingly revolves around making those models economically viable inside real organizations. Inference costs have become one of the largest constraints on enterprise AI adoption because every unnecessary token represents additional compute, latency, and expense. Companies reducing those costs without sacrificing performance are becoming increasingly attractive to both customers and investors.

Engram's approach places it squarely within the rapidly expanding AI infrastructure category, where optimization, memory architectures, continual learning, and deployment economics are becoming just as strategically important as model capabilities. The market is beginning to reward companies solving the operational realities of enterprise AI rather than simply pushing benchmark scores higher.

Commercial Validation

Early commercial traction provides additional context for the funding. Microsoft is piloting the technology inside Microsoft 365 while supporting model training through Microsoft Azure and Dapple. Notion is evaluating custom AI agents capable of understanding large workspaces without constantly starting over. Harvey is applying the technology within legal AI, where institutional knowledge often represents a firm's most valuable competitive advantage.

These are not random design partners. They represent environments where organizational memory directly influences productivity, decision quality, operating costs, and long-term competitive advantage. Enterprise software increasingly succeeds not because it knows everything, but because it remembers what matters.

What This Signals

There is a broader startup lesson hiding beneath the funding announcement. Engram did not raise $98M by chasing the loudest trend. The team focused on a problem many organizations had simply accepted as the cost of doing business with AI. Instead of asking how to build a bigger model, they asked why intelligent systems keep forgetting what they already learned.

That distinction matters because the AI industry has spent the past several years proving machines can reason. The next competitive advantage may belong to companies that teach those systems how to remember. If enterprise AI becomes less about constantly rereading information and more about accumulating institutional knowledge over time, memory could become one of the defining infrastructure layers of the next generation of AI.

---