An AI memory startup called Engram has raised $98 million, targeting two of the most expensive problems in modern AI development: model efficiency and token costs.
Every time an AI model processes text, it burns through what are called tokens, the basic units of language a model reads and generates. More tokens mean more compute. More compute means higher costs. Engram’s pitch is that smarter memory architecture can reduce how many tokens a model needs to use in the first place.
What Engram is actually building
Most large language models have a fundamental architectural limitation: they process information within a fixed context window. Once a conversation or document exceeds that window, earlier information falls out of view. Memory-focused startups are building systems that extend or augment this capability, giving models something closer to persistent, structured recall. Reducing redundant token processing is where the efficiency gains come in.
Engram’s $98 million raise has been reported as of June 23, 2026, though no details on round type, lead investors, or company valuation have been made public. No specific founders, executives, or backers have been named publicly. No product architecture, go-to-market strategy, or enterprise customers have been disclosed.
