This paper defines LM Software — a new category of software where applications are developed by configuring language model sessions with structured state, encoding logic as Prolog rules over exact arithmetic primitives, validating behavior interactively, snapshotting working state as deployable artifacts, and cloning those artifacts on demand as running instances. The language model is the runtime. A hierarchical knowledge base tree is the address space. Prolog is the programming language. Snapshots are the binaries. Usage improves the application without rebuilding it. LM Software is distinct from conventional software (compiled code, static behavior, developer-built), from AI tooling (LLMs wrapped in conventional orchestration code), and from conventional LLM applications (stateless models behind prompt templates). It is software developed by users through conversation, deployed as frozen state snapshots, and improved by usage through rule accumulation — where every session deposits reusable logic that makes future sessions cheaper and more capable. This paper introduces all necessary concepts from the VDR-LLM-Prolog architecture @HOWL-VDR-14-2026 so that no prior reading is required, then develops the theory and practice of LM Software through two complete worked examples.
Geoffrey Howland (Fri,) studied this question.