Engram builds a "learned memory" layer that sits underneath enterprise AI systems. The core problem it solves: today's AI models re-read the same documents and re-learn the same context on every single query, burning enormous amounts of tokens and money. Engram's models instead study an organization's files, workflows, and history in advance, compressing that knowledge into compact, reusable memories called engrams. On subsequent queries, those memories are recalled rather than recomputed — reducing token consumption by up to 100x while matching or outperforming frontier-model accuracy. The technology roots in a Stanford research technique ("Cartridges") that compresses long documents into small, reusable model memory. Buyers are enterprises running AI at scale in token-hungry domains: legal research, document workflows, agent automation. Early customers include Microsoft, Notion, and legal AI startup Harvey. The company was founded in October 2025, came out of Stanford's AI lab, and went from zero to $98M raised in under eight months.