Engram Launches With $98M for 'Learned Memory' AI

Engram emerged from stealth on June 23 with $98 million in fresh capital and an unusual pitch: instead of feeding an AI model more context, teach it to remember less. The round was backed by General Catalyst, Kleiner Perkins and Sequoia, with angel checks from former OpenAI co-founder Andrej Karpathy — who recently joined Anthropic — Wiz CEO Assaf Rappaport and Berkeley AI pioneer Pieter Abbeel. The eight-month-old, 13-person company says its models can match or outperform frontier labs on enterprise tasks while burning up to 100× fewer tokens.

The bet targets the single line item that has become a tax on every enterprise AI deployment: context. Today's models reread an organization's documents, wikis and tickets on every query, and each of those tokens costs money and latency. Engram instead trains a model to "study" a customer's knowledge once, ahead of time, and compress what it learns into a compact, reusable memory. "We do that studying once, ahead of time, training the model to compress everything it learns into a compact memory it can reuse on every query," said CTO Sabri Eyuboglu. The company's framing is vivid: a 70,000-word legal contract that occupies roughly 400 kilobytes on disk can balloon past 100 gigabytes inside a model's working memory — and Engram's job is to shrink that studying step down to between 1% and 10% of the usual token load.

The founding team leans heavily on academic pedigree. CEO Dan Biderman holds a PhD in computational neuroscience from Columbia and did postdoctoral work at Stanford under Chris Ré, who is a co-founder; Eyuboglu, also a Ré protégé, created the "Cartridges" document-compression method the product builds on. The roster includes Berkeley's Jessy Lin (ex-Meta FAIR), Cornell's Jack Morris and tenured Stanford statistician Scott Linderman. The name itself is borrowed from neuroscience: an engram is the physical trace a memory leaves in the brain — a fitting metaphor for a company trying to give models something closer to long-term recall than a giant scratchpad.

Crucially for a stealth launch, Engram arrives with paying interest rather than a demo. The company says it is already being tested inside Microsoft 365 — a relationship that includes GPU capacity via Azure — as well as inside custom agents at Notion and at legal-AI firm Harvey. "You've got this explosion of data, explosion of cost," said Kleiner Perkins partner Leigh Marie Braswell. "Engram comes in and basically maps out your organization and offers orders of magnitude cheaper output." The new money will go toward compute and hiring.

The launch lands in the middle of an industry-wide scramble over the economics of inference. As frontier labs trade token-price cuts to win enterprise workloads, a layer that quietly removes 90% or more of the tokens a query needs is attacking the same problem from the opposite direction. Whether a startup's specialized memory can keep pace with general-purpose models that grow cheaper and smarter every quarter is the open question — but $98 million and a client list of Microsoft, Notion and Harvey buys Engram a long runway to find out.