Key Takeaways
- • Engram co-founders Dan Biderman (neuroscience background) and Jessy Lin (Stanford PhD in computational cognitive science, 2007) are building an AI research firm focused on solving catastrophic forgetting in machine learning systems through memory and continual learning architectures.
- • Engram's core thesis positions memory and continual learning as "two sides of the same coin"—complementary mechanisms designed to enable AI systems to learn from new data without degrading performance on previously learned tasks, a critical limitation in current enterprise AI deployments.
- • The startup is developing what it calls "always-learning enterprise models" that can scale compute on user context, allowing AI systems to learn directly from proprietary organizational data while maintaining long-term retention and knowledge integration.
- • Sparse memory finetuning research cited in Engram's technical work demonstrates substantial performance preservation: while baseline models experience 89% F1 score degradation on NaturalQuestions when learning new knowledge, sparse memory approaches significantly reduce this forgetting penalty.
- • The company has attracted attention from Sequoia Capital, which featured the founders in a neolab discussion, indicating institutional venture interest in foundational AI research addressing persistent technical limitations in production language models.
Summary
Engram, a newly emerging AI research firm, is addressing one of the fundamental technical challenges limiting enterprise deployment of large language models: the inability of AI systems to continuously learn from new data without catastrophically forgetting previously acquired knowledge. Co-founders Dan Biderman and Jessy Lin, whose backgrounds span neuroscience and computational cognitive science respectively, have built the company around a focused thesis that memory and continual learning represent two sides of the same architectural problem.
The technical challenge Engram targets is well-documented in machine learning research. Current state-of-the-art language models suffer significant performance degradation when fine-tuned on new datasets—baseline models experience approximately 89% F1 score drops on benchmark tasks like NaturalQuestions when trained on new knowledge. This "catastrophic forgetting" phenomenon limits the practical applicability of AI systems in enterprise environments where models must adapt to new information while preserving existing capabilities.
Engram's approach centers on developing "always-learning enterprise models" that integrate memory mechanisms with continual learning architectures. The company's research demonstrates that sparse memory finetuning substantially reduces the forgetting penalty, allowing systems to acquire new knowledge while exhibiting meaningfully lower performance degradation on previously learned tasks. This technical innovation enables organizations to scale compute on their proprietary context—essentially allowing AI systems to learn directly from organizational data, documents, and workflows while maintaining long-term knowledge retention.
The founding team's interdisciplinary expertise positions them to bridge neuroscience-inspired AI architecture with practical enterprise machine learning. Lin's doctoral work in computational cognitive science provides theoretical grounding in how biological systems solve the learning and memory integration problem, while Biderman's neuroscience background informs the translation of these principles into scalable AI systems. The company's positioning around neolab research, as featured by Sequoia Capital, suggests institutional recognition that solving continual learning represents a significant unmet need in the AI infrastructure market, particularly for organizations seeking AI systems that improve over time rather than degrade in performance.
The technical challenge Engram targets is well-documented in machine learning research. Current state-of-the-art language models suffer significant performance degradation when fine-tuned on new datasets—baseline models experience approximately 89% F1 score drops on benchmark tasks like NaturalQuestions when trained on new knowledge. This "catastrophic forgetting" phenomenon limits the practical applicability of AI systems in enterprise environments where models must adapt to new information while preserving existing capabilities.
Engram's approach centers on developing "always-learning enterprise models" that integrate memory mechanisms with continual learning architectures. The company's research demonstrates that sparse memory finetuning substantially reduces the forgetting penalty, allowing systems to acquire new knowledge while exhibiting meaningfully lower performance degradation on previously learned tasks. This technical innovation enables organizations to scale compute on their proprietary context—essentially allowing AI systems to learn directly from organizational data, documents, and workflows while maintaining long-term knowledge retention.
The founding team's interdisciplinary expertise positions them to bridge neuroscience-inspired AI architecture with practical enterprise machine learning. Lin's doctoral work in computational cognitive science provides theoretical grounding in how biological systems solve the learning and memory integration problem, while Biderman's neuroscience background informs the translation of these principles into scalable AI systems. The company's positioning around neolab research, as featured by Sequoia Capital, suggests institutional recognition that solving continual learning represents a significant unmet need in the AI infrastructure market, particularly for organizations seeking AI systems that improve over time rather than degrade in performance.
Go deeper than the headline
Research every firm in this story, track the ones that matter and get the daily briefing — free for 7 days.