(Left to right) Jack Morris, Sabri Eyuboglu, Dan Biderman, Scott Linderman, and Engram’s Jesse Lin.
Provided by: Natalie Biderman
As American companies finally begin to crack down on unchecked AI use by developers, eight-month-old startup Engram sees a big opportunity in helping businesses reduce costs.
Engram announced Tuesday that it has raised $98 million from investors including General Catalyst, Kleiner Perkins, Sequoia, and OpenAI co-founder Andrej Karpathy, who recently joined Anthropic.
The startup, which bills itself as a “learned memory” for AI, says its models can remember an organization’s unique workflow and context to predict questions and give smarter responses with cheaper output. The company claims its models can use up to 100 times less tokens, the currency used to run AI queries, and can perform as well as or better than Frontier Labs.
New, more sophisticated AI models are proving to be more expensive than previous versions, challenging the traditional view that scaling up leads to lower costs.
“Data is exploding and costs are exploding,” said Lee Marie Braswell, a partner at Kleiner. “Engram comes in and basically plans the organization and delivers an order of magnitude cheaper output.”
In less than a year, the 13-employee company has amassed a client roster that includes: microsoftNotion and legal AI startup Harvey. Engram is derived from the neuroscience term for a memory trace in the brain, and the company plans to use the funding to support computing and human resources.
Engram co-founder and CEO Dan Biderman has a lifelong obsession with memory. He said it all started as a child when he tried to trick his grandmother, who had lost her memory, into remembering small facts about him and his siblings.
This ultimately led Biderman to pursue a PhD in computational neuroscience at Columbia University and then join Stanford University’s AI lab. While working at Stanford University, Biderman began to recognize what he called the “Genius Stranger Model,” the idea that AI is smart, but its memory is much more limited than it appears. At the same time, more context can overwhelm the model, requiring more research and reading, and higher costs.
Biderman acknowledges that Engram’s model is “absolutely not better” than models like OpenAI and Anthropic, but says it is better at specializing, sometimes at the expense of other features.
“We’re trying to go beyond this existing note-taking and build a layer of human intuition that isn’t present in current models,” Biderman said.
Spotlight: Fixing overspending on AI is a problem for OpenAI and Anthropic
