Memory CLAUDE.md sync

Whenever a markdown file under memory/ changes, the CLAUDE.md files above it must stay coherent. This skill keeps that hierarchy honest: every time you create, rewrite, rename, move, or delete a memory page, walk up the directory chain and patch each ancestor CLAUDE.md.

The claude-md-template skill will also fire because we are editing CLAUDE.md files — that is expected. This skill says what to change at the content level; claude-md-template ensures the structural spine stays intact.

Step 1 — Identify the chain

Skip read-only operations, non-markdown files (e.g. .obsidian/ configs), and anything outside memory/. If multiple memory pages change in one turn, run this skill once at the end of the turn and patch every affected ancestor in a single pass.

For each changed page P:

1. Walk up from P's directory toward memory/.
2. Collect every CLAUDE.md between P's directory and memory/ (inclusive of both). The workspace root CLAUDE.md is not in scope — it lists top-level workspace folders only and does not need updating when a single memory page changes.
3. Order the chain from closest-to-P (the immediate parent) to furthest (memory/CLAUDE.md).

The template's memory/ is flat: every page is a sibling of memory/CLAUDE.md, so for almost every page the chain is length 1 — memory/CLAUDE.md only. A leaf like memory/foo.md has just memory/CLAUDE.md above it. The multi-level walk-up below still applies if you grow memory/ into nested sub-folders (each with its own CLAUDE.md); for the default flat wiki, Step 2 is usually all you need.

Step 2 — Update the immediate parent

The immediate parent CLAUDE.md is where the changed page is listed by name. In the default flat wiki this is always memory/CLAUDE.md. Pick the right section by inspecting which catalog sections exist in that file:

• ## Projects — pages of type project (ongoing efforts, initiatives, areas of work).
• ## Concepts — pages of type concept or synthesis (themes, frameworks, mental models, analyses).
• ## Preferences — pages of type preference (durable preferences, conventions, standing instructions).
• ## Deep-Dive Navigation — links to child CLAUDE.md files. Only relevant when the changed entity is itself a navigation node (a sub-folder with its own CLAUDE.md), which is rare in the flat wiki.

Match the section to the page's frontmatter type field. Apply the right operation:

Read enough of the changed page to write an accurate description and pick keywords. Mirror the tone of the existing rows in the same table — they are usually one sentence ending in a period, with keywords as comma-separated lowercase tokens.

Step 3 — Walk further up

This step only applies if memory/ has nested sub-folders with their own CLAUDE.md files. In the default flat wiki the chain ends at memory/CLAUDE.md and you can skip ahead to Step 4.

For deeper hierarchies, the work in distant ancestors is lighter. You are not adding a row per leaf page in distant ancestors — that is precisely why intermediate CLAUDE.md files exist. You are checking whether the ancestor's description of the next step down has drifted.

For each remaining ancestor in the chain, in order:

1. Find the row in the ancestor's ## Deep-Dive Navigation table that points to the next step down the chain — e.g. [sub-area/CLAUDE.md](sub-area/CLAUDE.md) inside memory/CLAUDE.md.
2. Ask: does the changed page meaningfully shift the character of that subfolder (a new domain emerging, a new theme family, a removed pillar)?
3. If yes, refresh that row's description and keywords to reflect the broader picture.
4. If no, leave it alone.

Most leaf-page additions stop at the immediate parent. Only structural shifts in the folder's overall shape ripple further up.

Protections

• Never reorder existing rows or sections. Insertions go at the bottom of the relevant table unless an obvious alphabetical or thematic ordering is already in place — match the existing order if so.
• Never rewrite custom prose, operating rules, what-belongs-here sections, suggested-shape sections, or any non-table content unless the user asked you to.
• Preserve table column headers exactly — downstream tooling and the claude-md-template skill rely on Path | Description | Keywords.
• Use relative links rooted at the CLAUDE.md file being edited (foo.md from memory/CLAUDE.md, not absolute paths, not [[wikilinks]]).
• Standard Markdown only — [label](relative/path.md#anchor). Section anchors are #kebab-case-of-the-heading.

Step 4 — Verify links

Editing, moving, deleting, or renaming memory pages and rewriting nav rows is exactly when links break, so after patching every affected ancestor CLAUDE.md, run the link checker to confirm nothing broke.

1. Run the check. Prefer make check-links. If no makefile exists yet (pre-/setup), fall back to the direct command from the repo root: lychee --offline --include-fragments --config lychee.toml './**/*.md'.
2. Best-effort only. lychee is an optional dependency. If the command is not found, note that the link check was skipped (brew install lychee to enable it) and do not block the memory write on a missing tool.
3. Fix what this change caused. If the check reports broken links, fix the ones introduced by this sweep — a nav row pointing at a moved/renamed/deleted page, or a stale relative link — and re-run. For broken links you cannot confidently attribute or fix, report them rather than guessing.

Step 5 — Report

After the sweep, briefly tell the user which ancestors you touched and what changed — e.g. "Added a row for memory/foo.md to memory/CLAUDE.md under Projects." If you walked deeper ancestors, mention the ones you considered but left alone, so the reasoning is visible without forcing a diff review. Fold in the link-check outcome — clean, fixed (and what), skipped (lychee not installed), or unresolved links you flagged for the user.