/spec
MFM Spec
Durable project memory for coding agents. MFM Spec turns rough product intent into a validated graph of components, features, evaluations, and decisions that your own agent can read, evolve, and build from.
$ why it exists
Chat is a lousy source of truth. It disappears into scrollback, code outruns intent, and every new agent session wastes context rediscovering the same architecture.
MFM Spec makes the spec the durable artifact. The conversation that produces it is disposable; the validated graph remains portable, inspectable, and cheap for an agent to slice into context.
$ doctrine
MFM services do not replace your coding agent. They steer it. The agent brings the model; MFM supplies deterministic stores, validators, MCP tools, and skills: lines in the sand for the water to follow.
$ graph
componentsStructural owners of responsibility: where each durable part of the system lives.
featuresObservable behavior and the component capabilities each behavior needs.
evaluationsFeedback and judgment notes that record what was learned without rewriting intent by magic.
$ modes
localskillYour agent writes one manifest plus Markdown node files in the project, then validates the graph with the bundled linter.
hostedMCPA database-backed canonical spec with revision history, deterministic mutations, minimal-context reads, and portable export.
$ shape
my-system/
mfm-spec.yaml
system.md
components/
api.md
identity.md
features/
sign-in.md
evaluations/
sign-in-v1.mdThe directory layout is for humans. The authoritative shape is the graph: stable ids, typed node kinds, parent links, dependency edges, feature touches, and evaluation subjects.
$ integrity
Exactly one root component, named by the manifest.
Every parent, dependency, feature touch, and evaluation subject resolves.
Component and dependency graphs stay acyclic.
Responsibilities, intents, and summaries stay single-sentence and machine-sliceable.
Reasoning lives in the body, not only in frontmatter labels.
$ sibling offering
MFM Spec is one of the first MFM services. Atlas is another: a separate product and tooling discovery surface that serves grounded capability evidence when an agent needs to compare external tools.
explore Atlas