According to Dynamic Beating monitoring, AI memory startup Engram has announced the completion of a $98 million financing round, led by Sequoia Capital, Kleiner Perkins, General Catalyst, and recently joined by OpenAI co-founder Andrej Karpathy from Anthropic. Founded in October 2025, Engram, named after the term from neuroscience referring to "memory traces," currently has only 13 employees. The team describes itself as a large-scale model's "learned memory" layer, focusing on reducing enterprises' AI development and operational costs by improving model efficiency.
Addressing the pain point of high Token costs due to long-context, Engram claims that its model can reduce Token consumption by up to 100 times in specific workflows (i.e., requiring only one percent of the forefront model), while still achieving equivalent or better performance. Co-founder and CEO Dan Biderman noted that current large-scale models are facing an "unknown savant" dilemma, where the model itself is extremely smart but has limited memory. Although ultra-long contexts can hold more background information, each call requires re-retrieval and reading, leading to uncontrollable Token bills.
To cut costs, Engram has abandoned the development of all-purpose large models and instead opted to sacrifice some general abilities in exchange for specialized memory. This memory is dedicated to recording enterprise-specific workflows and contexts to predict issues and provide lower-cost responses. Currently, Microsoft, Notion, and legal AI startup Harvey have become early clients. This financing round will be mainly used for purchasing computing power and recruiting talent.
