PiPlan

Engram is a continuous-learning memory system that lets an LLM keep serving users while deciding what to store in RAG, what durable beliefs to internalize into model weights, and hot-loading those edits live in the background.