Alma organises knowledge into three layers — memories (facts), episodes (narrative summaries) and procedures (behavioural patterns) — assembles context in under 100ms with five-factor scoring (relevance, importance, confidence, recency, frequency), and links entities into concepts without LLM cost on read. The result: AI that actually remembers, not just chats.
Alma organizes knowledge into memories (discrete facts), episodes (narrative summaries), and procedures (behavioral patterns). Context assembly searches all three layers, scores by relevance/importance/confidence/recency/frequency, and builds the system prompt in under 100ms. Linked concepts add entity relationships without LLM cost.