memory-init

name	memory-init
description	Use this skill when bootstrapping memory in a freshly cloned repo or when upgrading from a pre-memory plugin version — to initialize the .ai-skills-memory/ skeleton in the current project (directory structure, .gitignore rules, learnings.md template, .committed/ allowlist). Idempotent — safe to re-run on a project that already has memory wired.
context	fork
argument-hint

/memory-init — Memory Skeleton Initializer

Create the .ai-skills-memory/ directory tree from plugin/memory/templates/. Idempotent. Layer model and write rules: plugin/docs/concepts/memory.md.

When to use

• After cloning a repo that uses ai-skills but has no .ai-skills-memory/ (first run on a fresh checkout)
• After upgrading the plugin to a version that adds new memory paths
• When /plugin-doctor reports "no memory tree present"

Not needed when running /ai-skills-init — that workflow runs memory-init as one of its steps.

Behavior

1. Detect current project root (<repo> = the cwd where the user runs the command).
2. Check for existing .ai-skills-memory/:
   • If missing → create
   • If present → no-op for existing files; create only what is missing (idempotent)
3. Create directory tree:

   <repo>/.ai-skills-memory/
   ├── .gitignore               (from plugin/memory/templates/ai-skills-memory.gitignore)
   ├── learnings.md             (empty header template)
   ├── runs.jsonl               (touch-create, append-only)
   ├── errors.log               (touch-create, append-only)
   ├── redactions.log           (touch-create, append-only)
   ├── sessions/                (per-session subdirs created by hooks)
   ├── ralph/                   (per-RALF-run subdirs)
   ├── pending-flush/           (PreCompact flush markers)
   ├── designs/                 (/feature-design output summaries)
   ├── env-reports/             (/env-analyze baselines)
   ├── refactor/                (/refactor logs)
   ├── migrate/                 (/migrate plans)
   ├── spikes/                  (/spike reports)
   ├── security-audits/         (/security-audit findings)
   ├── docs/                    (/docs-pack generation summaries)
   ├── knowledge-sync/          (/knowledge-sync per-run logs + run-lock)
   └── .committed/
       ├── README.md            (from plugin/memory/templates/committed-readme.md)
       ├── .allowlist-extensions.txt   (project-extension allowlist, empty by default)
       ├── conventions.md       (from plugin/memory/templates/conventions-schema.md, empty body)
       └── learnings.md         (committed-learnings template, empty body)
   

4. Append .ai-skills-memory/ rule to <repo>/.gitignore if not already present.
5. Print summary: created vs skipped paths.

Hard rules

• Never overwrite existing files. If a file exists, skip it and report "skipped: ".
• Never write to .committed/ paths outside the allowlist — pre-tool-use-committed-write.py hook enforces this; the skill must not bypass.
• PII filter applies to any seeded content per memory-discipline.md rule 2.

Failure modes

• No write permission in cwd: report error, suggest cd to a writable directory or run with elevated permissions.
• Existing .ai-skills-memory/ corrupted (bad JSON, bad allowlist): flag corrupted files but do not delete or rewrite them. Suggest user inspect and back up before re-running.

Memory writes

Layer	When	Shape
L4	On first init	.ai-skills-memory/init-summary.md — timestamp + version + paths created

Integration

• Templates source: ${CLAUDE_PLUGIN_ROOT}/memory/templates/ (7 files: ai-skills-memory.gitignore, committed-allowlist.txt, committed-readme.md, conventions-schema.md, eval-baseline.schema.json, learnings-schema.md, untrusted-content-wrapper.md). PII pattern file lives separately at ${CLAUDE_PLUGIN_ROOT}/hooks/scripts/pii-patterns.txt because the secret-scan + memory-curator hooks load it directly.
• Called by: /ai-skills-init (full bootstrap including memory), used standalone for upgrades
• Rules: memory-discipline (write rules per layer), memory-validation (entry validation)
• Hooks: pre-tool-use-committed-write.py validates .committed/ writes against allowlist
• Layer model: plugin/docs/concepts/memory.md documents the L0–L5 layers. This skill creates L3 dirs (sessions/, ralph/) and L4 files (learnings.md, runs.jsonl, etc.) inside .ai-skills-memory/.