Karpathy-Backed AI Memory Startup Engram Raises $98M to Build Continuous Learning Layer

AI startup Engram, founded in October 2024 and incubated at Stanford AI Lab, has emerged from stealth with $98 million in funding from top VCs including General Catalyst, Kleiner Perkins, and Sequoia, with AI luminaries Andrej Karpathy and Pieter Abbeel participating as investors/advisors. The 13-person team is valued at approximately $600 million.

Core Mission: Building a 'Memory Layer' for Continuous Learning

Engram addresses the 'forgetfulness' of large models in enterprise settings—models handle general knowledge but lack context on internal project decisions, historical discussions, etc., requiring repeated background in each conversation. Engram converts 'context' into model capability by offline training on enterprise data from GitHub, Slack, Notion, etc., rather than per-query retrieval. The goal is to update data from daily to hourly, eventually every minute, enabling continuous improvement during use.

Technical Highlights: Cartridges and Memory Compression

A key technology is Cartridges, led by CTO Sabri Eyuboglu (Stanford team, mentored by co-founder Chris Ré). It uses a 'self-study' process: generating synthetic dialogues around corpora and distilling 'learning traces' into compact caches. For a 70,000-word legal contract, traditional memory exceeds 100GB, while Cartridges reduces it to ~1/40 and boosts decoding throughput by 25x+. Engram also separates 'inference layer' from 'memory layer,' allowing real-time absorption of new data in seconds to hours without retraining.

Early Partnerships and Product

Engram's first product is an API for agents targeting large knowledge workspaces. Announced partners include Notion (Custom Agents understanding large workspaces), Harvey (law firm and enterprise knowledge scenarios), and Microsoft (customized agents in M365). These scenarios feature high knowledge density and complex context.

Team Background

CEO Dan Biderman holds a PhD in computational neuroscience from Columbia University and was a postdoc at Stanford AI Lab, researching memory and forgetting. The 13-person team hails from Stanford, Berkeley, and Cornell; several members turned down offers from Anthropic and Google. The company focuses on continuous learning, context compression, retrieval augmentation, LoRA, and synthetic data.

Industry Significance

Engram bets on a new scaling direction: investing compute in continuous learning of private context rather than merely scaling model size. Its goal is to transform AI from a 'genius stranger' into one that truly understands users and organizations, seen as a key piece toward AGI and even ASI.