Repo layout

A MemMini repo has a small number of directories with sharp roles. The golden rule: edit the source, never the build output.

context/                  intelligence layer, injected every turn (Policy lane)
  who.md                    behavioral rules (who you are)        [edit]
  how.md                    operating rules (how you work)        [edit]
  enforcement.md            routing / verifiable rules            [edit]
  reply-close.md            last-thing-you-read finishers         [edit]

wiki/                     knowledge layer, retrieved on demand   [edit]
  README.md                 the graph schema (frontmatter, edges)   (Knowledge lane)
  INDEX.md                  top-level hubs (compiled into CONTEXT.md)
  concepts/ patterns/
  contracts/ procedures/ plans/ services/
  platforms/{name}/         platform-scoped knowledge

skills/                   load-on-demand procedures (Knowledge lane) [edit]
  INDEX.md                  trigger lines
  {name}/SKILL.md           the procedure

scripts/                  the build + graph tooling
  (empty by contract)       generate a compiler from wiki/contracts/compile-declaration.md
  compile-wiki-registry.sh  builds .registry.json for slug lookup

.learnings/raw.md         capture log, append-only (Memory lane)
.coord/                   multi-agent coordination files (Coordination lane)

CONTEXT.md                GENERATED, the portable mind           [never hand-edit]
ENFORCEMENT.md            GENERATED, routing for your platform   [never hand-edit]

These four directories, context/, wiki/ + skills/, .learnings/, .coord/, are also the repo’s four information lanes: Policy, Knowledge, Memory, and Coordination. Each owns one kind of information and one lifecycle; see What is MemMini for what each one holds and must not become.

Canonical reference

The authoritative version of the source→artifact model and house rules for editing the repo is AGENTS.md. Its design rationale lives in the docs concepts and skills/distill/SKILL.md §The loop.