Karpathy Backs Engram: AI Memory Startup Aiming for Persistent Enterprise Knowledge

While many large‑model companies compete on longer context windows, stronger reasoning, and complex agent workflows, Engram focuses on a different question: can AI continuously learn from daily data, conversations, and experiences like humans do?

The startup has just announced a $98 million financing round backed by top venture firms (General Catalyst, Kleiner Perkins, Sequoia) and advisors such as Assaf Rappaport, Andrej Karpathy, and Pieter Abbeel. Both Karpathy and Jason Wei publicly congratulated the company.

Engram’s core problem is that current AI can answer questions, read code, and write documents, yet it remains unfamiliar with an organization’s internal knowledge—project decisions, historical trade‑offs, team discussions—because these details are rarely present in pre‑training data. Consequently, enterprises must repeatedly supply context, forcing the model to re‑understand each time, which is costly, error‑prone, and leads to forgetting after the conversation ends.

Engram argues that AI should not merely read context temporarily but truly learn it. Its positioning is to build a "memory layer" that enables continuous learning, allowing models to absorb internal company data (GitHub, Slack, Notion, documents, project records) in advance rather than retrieving and re‑reading on each query. This shifts computation forward compared to typical Retrieval‑Augmented Generation (RAG) or long‑context solutions.

The company bets on a new scaling direction: investing training compute into private, user‑specific contexts instead of extending chat windows. According to its description, Engram already has models learning company data daily and plans to increase the update frequency to hourly, eventually aiming for minute‑level updates. Its goal is a unified training algorithm that can ingest data of any scale or format and continuously improve model performance.

Engram’s first product is an API for agents that serves large shared‑knowledge workspaces. It has announced early collaborations with Notion (custom agents for large Notion workspaces), Harvey (focused on law‑firm and enterprise knowledge), and Microsoft (pilot in M365 for more efficient customized agents). These scenarios share high knowledge density, complex context, and difficulty for one‑shot retrieval, which Engram targets by enabling models to long‑term digest organizational information streams.

The research‑focused team behind Engram includes founders Dan Biderman, Sabri Eyuboglu, Jessy Lin, Jack Morris, and others whose backgrounds span continuous learning, context compression, retrieval‑enhanced models, LoRA, synthetic data, long‑context, and memory architectures. Their work centers on enabling models to learn from continuously changing data while avoiding forgetting and loss of control.

Engram’s strategic emphasis on continuous learning means it is not just adding an external memory module but building infrastructure—from training algorithms to system architecture to product experience—that supports long‑term learning. The key challenge remains proving that continuous learning can move from a research problem to a stable product that is reliable, controllable, auditable, and continues to deliver value after multiple update cycles.