| name | search |
| description | Cross-hub index search within the current workspace: manifest, closets, and the typed-edge graph, grouped by folder, with a built-in similarity pass for fuzzy/paraphrase matches. For cross-workspace search, use `/memex:cross-search`.
|
| argument-hint | <query> |
| disable-model-invocation | true |
Memex - Search
Wikilink rule: When referencing files in any markdown you write, use [[filename]] format.
Cross-hub search within the current workspace. Same retrieval surface as /memex:cross-search, scoped to one workspace.
Step 1: Resolve workspace and query
Run WORKSPACE_ROOT=$(pwd) && echo "$WORKSPACE_ROOT" via Bash.
Use $ARGUMENTS as the query. If empty, ask: "What do you want to search for across hubs in this workspace?"
If _MANIFEST.md is missing, tell the user "Memex isn't initialized in this workspace. Run /memex:init first." and stop.
Step 2: Search
Run ${CLAUDE_SKILL_DIR}/scripts/sources.py:
python3 "${CLAUDE_SKILL_DIR}/scripts/sources.py" search-local "<query>" --workspace "$WORKSPACE_ROOT"
The script greps _MANIFEST.md, every _CLOSETS.md and _CLOSETS-archive.md, and memory/.graph.md under the workspace. Output is grouped by folder.
Step 2b: Similarity supplement (built-in)
Always run this after grep, never instead of it:
python3 "${CLAUDE_SKILL_DIR}/scripts/semantic.py" query "<query>" --workspace "$WORKSPACE_ROOT"
Show hits under a Similarity matches section placed AFTER the grep results. Dedupe against grep hits: skip any result whose stem grep already surfaced.
Step 3: Format output
If results found:
- Show the script's grouped output verbatim (folder header, file:line:match lines).
- Summarize what cuts across hubs: if the same query hits multiple hub folders, call that out — that's the cross-hub signal the user came for.
- If a hit is in the Recently Archived section of a
_CLOSETS.md, suggest loading the corresponding _CLOSETS-archive.md for full typed-field context on that file.
If no results (grep and Step 2b's similarity pass both found nothing):
No matches for "<query>" in this workspace.
Either:
- The subject isn't wikilinked into the manifest or any closets file.
Run /memex:resummarize or /memex:reindex to refresh closets.
- The subject lives only inside content files. Open the relevant hub's
files directly (closets surface what's *in* the index, not raw content).
- The similarity pass (Step 2b) also found nothing — the subject likely
isn't indexed; run /memex:reindex.
Step 4: Suggested follow-ups
After results, append a one-line nudge based on what was found:
- Hits in multiple hubs → "Cross-hub query confirmed: subject lives in
<hub-a> and <hub-b>."
- Hits only in
_CLOSETS-archive.md → "Recent activity is archived. Load the archive for full context."
- Hits in
memory/_CLOSETS.md only → "Subject lives in Tier 1 (status / decisions / glossary / contacts)."
Gotchas
- Index-only search. Manifest and closets. Never inside content files. If a subject isn't wikilinked into a closet,
/memex:search won't see it. The fix is /memex:resummarize or /memex:reindex, not deeper grep.
- Case-insensitive substring, grep-only. Regex chars are escaped automatically; you can't use regex. The built-in similarity pass (Step 2b) is stdlib-only and catches morphological variants (fundraiser/fundraising) grep can't — but not true synonyms. Synonym-level matching engages automatically only in environments where
sentence-transformers is already present (e.g. a local dev machine); nothing to install, nothing to configure.
- Search is read-only. No writes, no closets refresh. To fix gaps surfaced by a search, run
/memex:reindex to refresh closets.
- Cross-workspace lives elsewhere. This skill scopes to one workspace. For "have I written about X in any of my workspaces", use
/memex:cross-search.
- Folder grouping uses path prefixes. Hits in
programs/ group together; hits in programs/subhub/ are a separate group. That's intentional — sub-hubs surface as independently as top-level hubs.
- Semantic cache is derived and disposable, embedding mode only.
memory/.semantic-cache/ only exists when sentence-transformers is present and the embedding engine ran — lexical mode writes nothing. Safe to delete any time; it rebuilds on the next semantic.py query. If the workspace is a git repo, consider adding it to .gitignore.
.graph.md hits are typed edges, not text matches. A hit in memory/.graph.md means a frontmatter field (people, supersedes, blocks, etc.) named the query term. Follow the edge to the named file(s) rather than treating the hit itself as the answer.
