AI memory startup focused on cutting token costs raises $98 million

With corporate America finally starting to crack down on untamed AI usage by developers, an 8-month-old startup called Engram sees a big business opportunity in helping companies save money.

Engram on Tuesday announced that it raised $98 million from investors including General Catalyst, Kleiner Perkins and Sequoia, as well as OpenAI co-founder Andrej Karpathy, who recently joined Anthropic.

The startup, which dubs itself the “learned memory” of AI, says its models can recall organization-specific workflows and context to anticipate questions and give smarter responses with cheaper output. The company claims its models can match or outperform frontier labs using up to 100 times fewer tokens, which are the currency for running AI queries.

New and more sophisticated AI models are proving pricier than previous iterations, challenging the conventional view that greater scale would lead to lower costs.

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

Less than a year after its founding, the 13-person company has accrued a client roster that includes Microsoft, Notion and legal AI startup Harvey. Engram, which comes from the neuroscience term for a trace of memory in the brain, plans to use the funding to support compute and talent.

Dan Biderman, Engram’s co-founder and CEO, has a lifelong obsession with memory. It started as a kid, he said, trying to trick his grandmother, who had lost her memory, into remembering little facts about him and his siblings.

That led Biderman to eventually pursue a PhD in computational neuroscience at Columbia University and later to join Stanford University’s AI lab. Working at Stanford, Biderman began to recognize what he calls the “genius stranger model” — the idea that AI is smart, but its memory is much more limited than it seems. At the same time, more context can overwhelm models, requiring more research and reading coupled with higher costs.

Biderman admits that Engram’s models aren’t “absolutely better” than those from the likes of OpenAI and Anthropic, but he says they excel at specializing — sometimes at the expense of other capabilities.

“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.

WATCH: The fix for overspending on AI is a problem for OpenAI and Anthropic