Engram raises $98 million to cut token costs

Three of the industry's most prominent venture firms backing a 13-person, eight-month-old startup at a reported $600 million valuation is less interesting than what they are betting on: that AI memory is splitting into two competing architectures. One is retrieval, where a system fetches relevant documents at query time and stuffs them into the context window, the basis of RAG and vector databases. The other, Engram's bet, is trained memory, where an organization's documents are compressed offline into the model's own weights, so it answers from what it has learned instead of re-reading source material on every call. For AI and ML teams evaluating token-cost fixes, this round is a signal that trained memory now has serious capital and research pedigree behind it, not just a product pitch.

Engram, founded in October 2025 by a team from Stanford's Hazy Research lab (CEO Dan Biderman, CTO Sabri Eyuboglu, plus co-founders Jessy Lin, Jack Morris, Scott Linderman and Chris Re), emerged from stealth on June 23 with $98 million in funding, according to the company's own announcement. Kleiner Perkins led the round, per its own investment post, with General Catalyst, Sequoia Capital, Amplify Partners and other investors also participating, alongside angel investors including Andrej Karpathy. Kleiner Perkins partner Leigh Marie Braswell describes today's models as "brilliant strangers" that "reread the same documents and relearn the same context" on every query. Engram's fix, descended from Eyuboglu's Cartridges research and Lin's sparse memory fine-tuning work, trains a small, reusable memory object offline on a customer's documents instead of reloading a retrieval cache at inference time. The result, Engram says, is models that match or beat frontier-model quality while using as little as 1 to 10 percent of the tokens, or up to 100 times fewer. Early testing partners include Microsoft, which is trialing the models inside Microsoft 365 and has committed Azure GPU capacity to the work, plus Notion and legal AI startup Harvey.

Baking knowledge into weights is not a drop-in swap for RAG. Biderman explained on Sequoia Capital's Training Data podcast that the approach requires white box access to model weights, so it works most easily on open-weight models; applying it to closed frontier models means partnering directly with the model provider, which is effectively the arrangement Engram has with Microsoft. That is a materially different integration path than adding a vector database. Trained memory also carries a staleness problem retrieval does not have: when source documents change, the compressed memory has to be retrained, while a vector index just gets re-indexed. And the up-to-100x figure so far comes only from Engram; independent analysis site The Deep Feed points out it has no third-party benchmark yet, so it is worth treating as a vendor claim pending outside verification rather than a settled number.

The competitive field is moving fast. The Deep Feed reports that rival Supermemory open-sourced its own memory engine about a week before Engram's announcement, a reminder that memory as a concept is commoditizing even as Engram bets its pipeline for turning a corpus into reliable memory, not the idea of memory itself, is the defensible part. Worth tracking over the next few quarters: whether Microsoft, Notion and Harvey convert their pilots into disclosed production savings, whether an independent party benchmarks the token-reduction claim, and whether frontier labs respond by building comparable memory features directly into their own model APIs.