Engram

Engram builds AI models that learn an organization’s knowledge before answering questions. Its first product is an agent API for large shared workspaces, designed to help agents reuse knowledge rather than repeatedly gather the same context. Engram starts with pretrained models and trains them on customer-specific information. Its research with Harvey combines knowledge stored in model parameters, structured notes, and document search, allowing agents to use learned memory alongside the underlying files.

Founded in 2025, Engram’s cofounders are CEO Dan Biderman, CTO Sabri Eyuboglu, Jessy Lin, Jack Morris, Scott Linderman, and Chris Ré. Its technical foundations include Cartridges, academic work by Eyuboglu, Ré, and collaborators that predates the company. Cartridges uses self-study on synthetic conversations to train compact, reusable attention caches for a document collection, spreading the preparation cost across subsequent queries instead of processing the full collection each time.

Engram publicly launched in 2026 with $98 million in funding. Its early partnerships connect organizational memory to existing software: it is developing Custom Agents for large Notion workspaces and models with Harvey that can find precedents across a law firm’s client matters. Its Microsoft 365 work remains a pilot for customized enterprise agents. These projects focus on repeated work within a shared knowledge base, where prior study can reduce the context an agent must gather for each task.