Jack Morris is an AI researcher and cofounder of Engram, which develops language models that acquire durable knowledge about the organizations using them. His work tackles a central limitation of contemporary AI: models can reason impressively about public information yet struggle to learn private documents, specialized skills, and the accumulated context of individual users.
Morris began researching natural-language processing as an undergraduate at the University of Virginia, where he helped create TextAttack, an open-source toolkit for adversarial attacks, adversarial training, and data augmentation. He became a Google Brain Resident before pursuing a doctorate at Cornell University, completed in 2025. There he worked with Alexander Rush and Vitaly Shmatikov, received an NSF Graduate Research Fellowship, and spent part of his PhD as a student researcher at Meta’s Fundamental AI Research group. He publishes academic research as John X. Morris.
His doctoral research exposed a consequential privacy problem in retrieval systems. Through embedding inversion, Morris and collaborators reconstructed text from numerical embeddings, recovering 92% of tested 32-token inputs exactly under their experimental conditions and identifying sensitive information in clinical text. The implication: embeddings cannot safely be assumed anonymous simply because their contents are unreadable to humans.
With Rush, Morris subsequently developed contextual document embeddings, which adjust a document’s representation according to surrounding documents. This helps retrieval systems distinguish material that looks similar in a general-purpose embedding space but differs meaningfully within a specialized collection. His work on language-model memorization also estimated 3.6 bits per parameter for the GPT-style models studied, giving practical definition to the finite capacity available for specialized knowledge.
Weight-based model memory: Information repeatedly inserted into prompts remains expensive to process and can become harder to use as context grows. Morris argues that adapting model parameters can make organizational knowledge persist across requests, while retrieval remains useful for fresh or frequently changing information. His analysis of prompting, retrieval, and model adaptation frames these as complementary tradeoffs between training-time investment and inference-time cost.
Learning without catastrophic forgetting: Training directly on a narrow document collection can produce nearly perfect memorization while damaging a model’s ability to answer unfamiliar questions. Morris emphasizes synthetic questions, varied training examples, and targeted parameter updates that preserve existing capabilities while helping models generalize.
TinyLoRA: In research on extremely small adapters, Morris and collaborators showed that reinforcement learning can improve mathematical reasoning with remarkably few trainable parameters, including experiments involving just one. The finding distinguishes compact behavioral adaptation from the more demanding task of storing substantial new factual knowledge.
Scaling compute on context: Morris’s approach to specialized model learning applies substantial computation to a fixed corpus through distillation, synthetic continued pretraining, and progressively harder self-improvement loops. The challenge is to keep deepening a model’s understanding after an initial synthetic training set stops yielding improvements.
A model can know the public literature yet lack depth in your domain. Learning from private context requires more than fitting documents: it requires training tasks that keep improving with the model.
A model can read your documents without learning to use them. Jack Morris explores what it takes to turn private knowledge into lasting capabilities—and where retrieval still belongs.
