$98 million went to a 13-person company on June 23, 2026. That ratio alone signals how much the market is betting on a specific architectural idea: that memory and reasoning in AI systems should not live in the same place.
Engram, which emerged from stealth Monday, disclosed the raise to CNBC alongside a roughly $600 million valuation. General Catalyst, Kleiner Perkins, and Sequoia backed the round, with participation from Andrej Karpathy, the former Tesla AI director and OpenAI founding member whose personal investments have become a closely watched signal in the AI infrastructure space.
The company’s core claim is architectural. Most production AI systems today use in-context memory, meaning they re-read the same documents, chat histories, and knowledge bases at the start of every session. That approach is reliable and interpretable, but it is expensive: every token processed is a token billed. Engram separates the reasoning layer from the memory layer entirely, allowing an enterprise or personal model to absorb new information through what it calls online continual learning, updating in seconds to hours without a full retraining pass and without losing what it already knew.
If the token-cost claim holds, the enterprise economics would be significant. A 100-times reduction in tokens consumed does not mean a 100-times reduction in total cost, because inference spend is only one component of a deployment budget, but it would meaningfully change the math on high-frequency, context-heavy workloads like legal review, code assistance over large repositories, and persistent customer-service agents.
The 100x figure and the “match or outperform frontier labs” assertion are Engram’s own, stated in its launch materials. CNBC’s reporting does not cite independent benchmark results, and Engram has not disclosed which models it compared against or on which tasks. The absence of third-party evaluation is standard for a stealth-to-launch announcement, but builders evaluating this architecture should treat both claims as hypotheses pending external replication.
The launch partnership with Microsoft, covering testing inside Microsoft 365, gives Engram access to one of the largest enterprise software distribution surfaces available. Notion and Harvey, the legal AI startup, are also listed among early clients. Founded less than a year ago, the company has moved from zero to $600 million in estimated value and a named Microsoft partnership in under twelve months.
The Karpathy participation is worth noting separately from the headline investors. His track record of backing technically credible infrastructure bets, and his public commentary on the inefficiencies of attention-based context handling, suggests he finds the continual-learning architecture genuinely plausible rather than merely fundable.
The structural question for any memory-layer startup is catastrophic forgetting: the well-documented tendency of neural networks to overwrite prior knowledge when trained on new data. Engram explicitly claims its architecture avoids this problem. That claim is testable, and the company’s credibility over the next twelve months will depend on whether enterprise deployments confirm it at scale.
For AI engineers and product teams currently budgeting for context-heavy deployments, Engram’s Microsoft 365 pilot is worth tracking as a real-world data point on whether the token-cost claims survive contact with production workloads.
Reporting by CNBC, published June 23, 2026.
