Engram is Building AI that Learns from You & Deeply Understands Your Work

Engram, scaling compute on your context, led by Dan Biderman, Scott Linderman, Jessy Lin, Sabri Eyuboglu, and the team, has raised $98M from General Catalyst, Kleiner Perkins, Sequoia, Factory, Modern, Amplify, Neo, SV Angel, and others, including advisors and investors Assaf Rappaport, Andrej Karpathy, and Pieter Abbeel.

The Missing Context for AI.

Most AI models are trained primarily on publicly available information, giving them broad knowledge across topics such as open-source code, documentation, and online content. However, they often lack an understanding of the proprietary knowledge, workflows, and context that define an individual's or organization's day-to-day work.

Bridging this gap is essential to making AI more useful, enabling models to move beyond general knowledge and operate with a deeper understanding of the information, processes, and expertise that drive real-world work.

A Different Approach to AI Training.

The company is pursuing a fundamentally different AI training strategy by starting with strong pre-trained models and focusing additional training on an organization's proprietary knowledge rather than relying primarily on public internet data. This approach is designed to help models develop a deeper understanding of company-specific context, relationships, workflows, and historical knowledge.

Through internal deployment and partnerships, the models have been trained on sources such as GitHub repositories, Slack conversations, and Notion workspaces, enabling them to understand ongoing projects, retain institutional knowledge, and surface insights that might otherwise be overlooked. Because much of this context is already internalized, the models can perform many tasks more efficiently without repeatedly retrieving the same information.

The long-term objective is to build a continual learning system that continuously incorporates new organizational knowledge into the model, allowing it to improve over time. The company believes this represents a new scaling direction for AI, using compute to deeply internalize proprietary context, enabling models to function more like knowledgeable teammates than general-purpose assistants.

Early Product and Partnerships.

The company’s first product is an API designed for agents that learn from large, shared knowledge workspaces. It is being developed with early partners that manage rich, high-value organizational data and have been among the earliest adopters of AI in their respective domains.

With Notion, the work focuses on building custom agents that can understand and reason over large Notion workspaces. With Harvey, the system is being adapted to internalize firm-wide legal knowledge and enable fast retrieval of relevant precedents across client matters. With Microsoft, Engram models are being piloted within Microsoft 365 to power cost-efficient, customized enterprise agents at scale.

Across these deployments, the goal is consistent: enabling AI systems to operate directly on deeply contextual organizational knowledge rather than generic external data.

A New Interaction Paradigm.

Future AI systems are expected to operate with far richer and more continuous context than today’s models, generating and processing trillions of tokens daily as they adapt to rapidly changing environments.

Rather than relying solely on static training data, these models will continuously incorporate new information from user interactions, effectively learning from ongoing workflows. Over time, this creates a persistent, evolving system shaped by the context it is exposed to, making the model increasingly personalized to the individuals and organizations that use it.