Engram raises $98M for AI memory
Engram left stealth with $98M to build a learned memory layer that, it says, lets enterprise AI match frontier models using up to 100x fewer tokens.
Engram came out of stealth on June 23 with $98 million in funding to build what it calls a learned memory layer for enterprise AI. The round was led by General Catalyst, Kleiner Perkins and Sequoia Capital, with angel backing from Wiz CEO Assaf Rappaport, OpenAI co-founder Andrej Karpathy and Berkeley researcher Pieter Abbeel. Reporting puts the valuation at $600 million on a team of about 13.
The problem Engram targets is concrete. Most AI tools re-read the same company documents and re-learn the same context on every single query, which burns tokens and money. Engram instead trains a model on an organization’s documents ahead of time, compressing what it learns into a compact memory it reuses on each request. The company claims this lets its models match or beat frontier models while using “1 to 10 percent of the tokens,” or up to 100x fewer. It names Microsoft, Notion and the legal AI firm Harvey as early partners.
What it means for you
The cost story is the part worth your attention. Tokens are what you pay for, and re-feeding the same context repeatedly is pure waste, so anything that cuts it goes straight to your bill. The caution: the 100x figure is Engram’s own, measured on its terms, and a new vendor holding your organizational memory is its own lock-in to weigh. This is the same pressure pushing inference prices down across the board and chipmakers to rethink their stacks. Before you switch anything, judge the tool against your real workload, not the launch numbers.