Loop & Graph: why chatbots fail without a memory

A chatbot fails in a company for a boring reason. The useful facts live in people’s heads, in old threads, and in tools that do not talk to each other. The model is asked to be brilliant on an empty table.

Loop & Graph is the opposite setup. A loop is a unit of repeating work: intake, fulfillment, follow-through, learning, governance. A graph is the shared memory of facts, rules, systems, people, and open questions. Every turn of an agent has four phases: pack, act, extract, ingest.

Pack pulls the smallest set of accepted facts that bear on this job. Act does the work with those facts in view, and asks if a required fact is missing. Extract lifts new facts and labels them proposed, accepted, or ask. Ingest writes them back. Only a human moves a material fact to accepted.

That protocol is how a one-person practice can operate like a firm, and how a client company can keep an agent honest after the installer leaves. The graph compounds. Chat history decays.