MD-OS (Artificial Prefrontal Cortex) v5.0 is a Markdown-native Operating Filesystem for persistent AI agents and robotic systems.
The paradigm is Operational Context as Filesystem: the agent’s operating context is externalized into readable, auditable, reconstructible, and actionable files.
Natural language is part of the operating surface, but it is not left as ephemeral prompt text. It is stabilized into files that can be inspected, corrected, replayed, and used by different hosts.
In familiar LLM terms, MD-OS is a distributed, persistent prompt/control plane: the prompt is spread across Markdown knowledge, natural-language programs, runtime state, append-only memory, connector contracts, and deterministic builders instead of living only inside one model context window.
The main program is distributed across:
Natural language acts as:
Markdown behaves as high-level batch semantics:
Natural-language programs live under:
md-os/ops/programs/
They compile into:
md-os/ops/compiled/programs.json
md-os/ops/compiled/programs.md
The agent is therefore not only prompted. It is programmed through stable textual artifacts that can be compiled, inspected, corrected, and replayed.
Natural-language programs should decompose complex objectives into small agentic tasks, following the compositional philosophy of UNIX:
small bounded task -> verified file artifact -> small bounded task
Each task must expose a file-level contract for declared inputs, outputs, permissions, budget, verification, and failure state. The natural-language layer describes semantic intent and constraints; deterministic builders and connectors implement repeatable mechanics. No stage may depend on private chat context as its only input, and no unverified output may silently become the next stage’s truth.