Engram raises $98 million to help companies lower AI token costs with smarter memory models

AI is getting smarter, but for companies paying to run it, it’s getting a lot more expensive. As businesses look for ways to control rising token costs and curb unchecked AI use by developers, AI startup Engram says it has found an opening.

The 8-month-old company, which builds memory models for enterprise AI systems, announced Tuesday that it has raised $98 million from investors including General Catalyst, Kleiner Perkins, Sequoia, and OpenAI co-founder Andrej Karpathy.

Engram is betting that one of the biggest problems in enterprise AI isn’t model quality alone, but the cost of giving those models enough context to be useful. Its pitch is that by helping AI systems remember company-specific workflows, internal knowledge, and recurring tasks more efficiently, businesses can get smarter responses without burning through as many tokens.

Engram is going after a problem that is starting to hit large companies where it hurts most: the cost of running AI at scale. Newer models may be more capable than those that came before, but they are hardly cheap. The more context they ingest, the more tokens they burn, and the higher the bill climbs. For businesses trying to roll out AI tools across engineering, legal, and internal knowledge work, that math is becoming harder to ignore.

Engram Wants to Give Enterprise AI a Better Memory at a Lower Cost

Engram’s answer is a “learned memory” layer for AI systems. The idea is to give models a better way to remember how a company works, what its teams care about, and how recurring tasks are handled, without forcing them to repeatedly process huge amounts of context from scratch. The startup says that it lets its models anticipate questions, retrieve the right organizational knowledge, and deliver answers using far fewer tokens than general-purpose frontier models.

The company claims its systems can match or outperform top lab models on some enterprise tasks while using up to 100 times fewer tokens. That is a striking promise in a market where token efficiency is starting to matter almost as much as raw model performance.

“You’ve got this explosion of data, explosion of cost,” Leigh Marie Braswell, a partner at Kleiner Perkins, said in an interview. “Engram comes in and basically maps out your organization and offers orders of magnitude cheaper output.”

That pitch has already attracted a surprisingly heavyweight customer list for a startup that is less than a year old. Engram says its clients include Microsoft, Notion, and legal AI startup Harvey. The company has 13 employees and plans to use the new funding for compute and hiring.

The name Engram comes from neuroscience, where it refers to a trace of memory in the brain. That is not branding by accident. Dan Biderman, the company’s co-founder and CEO, has spent years thinking about memory, first on a deeply personal level and later as an academic obsession.

Biderman said his fascination started in childhood, when his grandmother began losing her memory. He would try to prompt her to remember small details about him and his siblings, an experience that stayed with him and later shaped his academic path. He went on to pursue a PhD in computational neuroscience at Columbia and later joined Stanford’s AI lab.

At Stanford, Biderman said he began to see a gap between how powerful AI models appear and how limited they can be once they are deployed in real organizational settings. He calls it the “genius stranger model.” A model may be smart in the abstract, but it still behaves like an outsider if it lacks durable context about the people, processes, and patterns inside a company. Stuffing more context into the prompt can help, but that tends to raise costs and create its own performance tradeoffs.

Biderman is not arguing that Engram has built a better frontier model than OpenAI or Anthropic. His claim is narrower and more practical. Engram’s models, he said, are built to specialize. That means trading some general-purpose breadth for the ability to learn how a particular organization operates and respond more efficiently inside that environment.

“We’re trying to go beyond this existing notetaking and build this layer of intuition that humans have, and current models don’t,” Biderman said.

That idea lands at a moment when companies are becoming more disciplined about how AI is used internally. Over the past year, developers have embraced coding assistants and AI tools with the kind of enthusiasm that tends to show up long before finance teams see the invoice. What comes next is likely a more sober phase of enterprise AI adoption, one where buyers still want the productivity gains but are no longer willing to tolerate open-ended spending.

If Engram can deliver on its promise, it may have found a timely opening in the market: helping companies retain the benefits of AI without inheriting the full cost structure of frontier-model usage.