Memory and Continual Learning: Engram's Dan Biderman and Jessy Lin

AI models currently rely on RAG and context engineering, which are computationally expensive and fail to capture the deep, intuitive understanding required for complex, evolving knowledge work. Engram addresses this by developing "always training" models that utilize adapter fine-tuning to internalize team-specific data directly into model weights. This approach reduces inference token consumption by orders of magnitude while enabling agents to learn from private, bespoke workflows that generic frontier models cannot master. By treating memory as a fundamental component of model architecture rather than an external database, these systems move toward a future where personalized AI agents evolve alongside their users. This shift from static retrieval to continuous, weight-based learning mimics biological memory, allowing models to synthesize complex associations and operate with the efficiency of an experienced employee who deeply understands company initiatives and internal processes.