Summary

• Web Config tab now shows indexing.memory_dirs as a categorized, editable widget (User / Claude projects / Claude plans / Codex collapsible groups) with per-item remove + reindex + inline add, replacing the previous 30-path comma-joined read-only line.
• Each dir gets a "(N chunks)" / "(not indexed)" badge fed by a new GET /api/memory-dirs/status endpoint, so users can see at a glance which dirs still need a manual reindex.
• 7 new unit tests for the aggregation helper; 18 new i18n keys with en/ko parity.

Why

After v0.1.12 (#292) moved provider memory dirs to opt-in explicit entries, a typical post-migration install ends up with ~25 ~/.claude/projects/*/memory paths added to config.json. Two unrelated code paths then failed the same user expectation at once:

• settings-config.js rendered memory_dirs via _READONLY_FIELDS.indexing + val.join(', ') — a flat, uneditable line. POST/POST /api/memory-dirs/{add,remove} existed as APIs but the UI never surfaced them.
• lifespan.app_lifespan only started the FileWatcher (indexing/watcher.py:85-96), which schedules Observer listeners but never walks existing files. Files that pre-existed the migration were invisible until the user manually ran mem_index. So "I added the dirs, why isn't my memory there?" was structurally unanswerable from the UI alone.

Both problems share the same trigger but are independent bugs — one UI, one startup lifecycle. Bundling them keeps the user-facing story coherent (the Config tab now tells the truth: what's watched, how to prune it, how to kick a reindex — and which dirs are still empty).

Why pull-based instead of gated startup scan

The first commit on this branch (f611d47) tried to close the indexing gap with a background asyncio.Task that scanned every memory_dir with zero chunks at startup. Local smoke on a real post-migration config (29 dirs, ~2000 .md files) made that design untenable:

• ONNX embedding the full backlog ran for minutes and left the UI unresponsive.
• Every per-file upsert_chunks failure (e.g., pre-existing schema drift in the user's DB) flooded the console with identical ERROR lines — ~200 copies of the same message.
• Users had no way to see what was being indexed or how long it would take.

The second commit (918c105) pivots: no automatic scan. Startup is back to "watcher only". The widget fetches per-dir status and displays (N chunks) or (not indexed) badges; the user clicks [↻] when they want to pay the reindex cost. This makes the "needs work" state visible without imposing a blind cost.

Changes

Part A — memory_dirs becomes an editable widget

• settings-config.js: remove memory_dirs from _READONLY_FIELDS.indexing. Introduce _NO_RESET_FIELDS to suppress the ↺ button for widgets that persist each action immediately (reset-to-default is semantically meaningless here). Register _buildMemoryDirsWidget in _CONFIG_CUSTOM_WIDGETS.
• The widget groups paths client-side via _categorizeMemoryDir, mirroring _detect_provider_dirs in config.py. Categories: user, claude-memory, claude-plans, codex. user is open by default; others start collapsed (Claude projects can be 20+ entries).
• Each row has [↻] (POST /api/index) and [✕] (POST /api/memory-dirs/remove). Each group summary has [↻] (loops over category). A top-level [↻ Reindex all] calls POST /api/reindex. Bottom input calls POST /api/memory-dirs/add.
• The widget lives outside the per-section Save/dirty flow by design: memory_dirs mutations carry side effects (watcher schedule, DB chunk state) that must land atomically on the server, not batched via PATCH /api/config.
• Confirmation dialog for removal uses showConfirm; all toasts go through t('toast.memory_dir.*', {...}). Covered by the existing i18n regression guards in test_i18n.py (no template-literal toasts, no English-literal titles).
• style.css: .memory-dirs-* classes. Claude-memory group gets max-height: 220px; overflow-y: auto so 25+ entries don't blow out the card.

Part B — per-dir index status

• indexing/engine.py: new memory_dir_stats(storage, memory_dirs) helper returning [{path, chunk_count, source_file_count, exists}] for each configured dir. Aggregates via a single storage.get_source_files_with_counts() call bucketed by normalised-path prefix — avoids N LIKE queries for large dir lists.
• web/routes/system.py: new GET /api/memory-dirs/status endpoint returning the helper's output.
• settings-config.js widget: fetches status after first paint (non-blocking); re-fetches after add / remove / reindex. Per-item badge renders {count} chunks, greys out the path text for 0-chunk dirs, and strikes through + italicises it for dirs missing on disk. Group summary shows {files} files · {chunks} chunks aggregate.
• 4 new i18n keys (settings.memory_dirs.status_{empty,missing,chunks,group}) with Korean translations and placeholder parity.

Not changed (out of scope)

• Startup auto-scan: attempted in commit 1, reverted in commit 2. Pull-based is what ended up shipping.
• _canonical_provider_dirs() scope narrowing — the migration already ran; retroactively trimming would surprise users who rely on the dirs.
• Promoting memory_dirs management to its own top-level Sources tab — discussed in review; worth a follow-up once the widget's current layout is proven in practice.
• Cloud-sync mount handling (GoogleDrive etc.) — separate tracked issue.
• _resolve_namespace behaviour where parent == memory_dir_root + enable_auto_ns=true collapses to default — pre-existing design gap, will file a follow-up issue.

Tests

• New (7): TestMemoryDirStats in test_indexing_engine.py — empty input, missing dirs, empty dirs, aggregation, per-dir bucketing, nested-file coverage, input-order preservation.
• Existing: full pytest -m 'not ollama' suite green (1845 pass, 46 ollama-deselected). Relevant sub-suites exercised specifically: test_web_routes.py, test_web_hot_reload.py, test_web_routes_extended.py, test_web_exclude_guard.py, test_i18n.py. No regressions.
• Lint: uv run ruff check + uv run ruff format --check pass. uv run mypy clean on the four touched Python files.
• JS: node -c settings-config.js clean (no syntax errors).
• Manual smoke (see Test plan): mm web against a post-migration 29-dir config — widget renders all 4 categories correctly, per-dir badges populate after initial status fetch, [↻] triggers reindex and badges refresh.

Test plan

• Restart mm web on a post-v0.1.12 config.json (many Claude-project dirs). Confirm Config tab → Indexing → Memory Dirs renders 4 collapsible groups with correct counts; User open, others collapsed.
• Within ~200ms of paint, each dir item shows a (N chunks) or (not indexed) badge. Group summaries show {files} files · {chunks} chunks.
• Click [✕] on one Claude-project entry → confirm dialog → accept → verify entry disappears + ~/.memtomem/config.json no longer contains it.
• Use [+ Add] to add /tmp/test-memdir-$$ → verify new User-group entry + badge shows (not indexed) (chunk_count=0).
• Click per-row [↻] on an empty dir → toast "Reindexing …" → on completion the badge flips to {N} chunks.
• Click group-level [↻] on claude-projects → sequential reindex across all dirs in the group; badges refresh after each.
• Click top-level [↻ Reindex all] → POST /api/reindex; badges refresh with aggregated totals.
• Check logs: no "Initial scan: indexing N memory_dir(s)" message at startup (confirms pull-based design).
• Delete a dir from disk (rm -rf /tmp/test-memdir-$$) without removing it from config → reload page → badge shows missing (strikethrough + italic).
• Sanity-check Korean locale renders the 4 category labels and all badge messages correctly.

🤖 Generated with Claude Code