kendreaditya/memex

Memex

Single-query unified search across the user's personal data — Logseq, Drive, Gmail, iMessage, WhatsApp. This skill is a playbook: it tells you which CLI to invoke for each source, how to fan them out in parallel, and how to present the aggregated results. No dedicated CLI — the backing tools already exist.

When to use

Use memex for questions about the user's own life/history/records:

• "what did I write about vraja?"
• "find everything mentioning the madhuri doc"
• "when did we first talk about the breakup"
• "pull up anything about my birthday"
• "search my life for 'kirtan'"

NOT for: the public internet (use forage for that), the user's code repos (use Grep/github), or the user's money (use monarch-money).

How it works

For each query, search all available sources in parallel — issue one Bash call per source in a single message. Then synthesize into one grouped result.

Source selection

Default: every available source. Narrow only if the user asks ("just my email", "only logseq").

| Source | Backend | When it's the strongest signal | | ----------- | ------------------------------------ | ----------------------------------------------------------------------- | | logseq | logseq search | Thoughts, reflections, journals — the user's own written processing | | drive | gog drive search | Long-form docs, planning notes, spreadsheets, media | | gmail | gog gmail search | Formal correspondence, receipts, calendar invites, shared Google Docs | | imessage | direct sqlite3 on chat.db | Real-time conversations with family / close friends (iMessage + SMS) | | whatsapp | wacli messages search | Group chats, overseas contacts, anything not in iMessage |

Availability check before running: which logseq gog wacli; test -r ~/Library/Messages/chat.db && echo imsg-ok.

Search commands

Replace QUERY with the user's search term. Always request JSON so synthesis is structured.

logseq

logseq search "QUERY" --format json --limit 10

Returns blocks, pages-content, pages. Strip the PFTS $pfts_2lqh>...<pfts_2lqh$ markers from snippets. No dates in raw results — follow up with logseq backlinks "<page>" --with-dates or logseq block <uuid> for journal day.

Google Drive

gog drive search "QUERY" --json --results-only

Fields: name, modifiedTime, mimeType, webViewLink, id. Use modifiedTime[:10] as the date.

Gmail

gog gmail search "QUERY" --json --results-only
# with date filter — use Gmail's native syntax (faster than client-side filtering):
gog gmail search "QUERY after:2025/08/01 before:2025/12/31" --json --results-only

iMessage

No built-in search in imsg. Query chat.db directly (read-only). Needs Full Disk Access for the invoking terminal.

sqlite3 -readonly ~/Library/Messages/chat.db <<'SQL'
SELECT
  datetime(m.date/1000000000 + 978307200, 'unixepoch', 'localtime') AS when_local,
  CASE WHEN m.is_from_me = 1 THEN 'me' ELSE COALESCE(h.id, '?') END AS who,
  REPLACE(m.text, CHAR(10), ' ') AS text
FROM message m
LEFT JOIN handle h ON m.handle_id = h.ROWID
WHERE m.text LIKE '%QUERY%' COLLATE NOCASE
ORDER BY m.date DESC
LIMIT 10;
SQL

Notes: (1) Apple stores m.date as nanoseconds since 2001-01-01 — the /1e9 + 978307200 converts to Unix epoch. (2) m.text is NULL for rich-content-only messages (stored in attributedBody BLOB) — those won't match. (3) Add AND m.date > (strftime('%s','YYYY-MM-DD') - 978307200) * 1000000000 for date filtering.

WhatsApp

wacli messages search "QUERY" --limit 10
wacli messages search "QUERY" --after 2025-08-01 --before 2025-12-31 --limit 10

First-time setup: wacli auth (QR login + initial sync). If the user hasn't run this, the source will return empty — tell them to run it.

Orchestration — parallel fan-out

Send a single assistant message with multiple Bash tool calls in parallel (one per available source). Example for query "madhuri":

[Bash: logseq search "madhuri" --format json --limit 10]
[Bash: gog drive search "madhuri" --json --results-only]
[Bash: gog gmail search "madhuri" --json --results-only]
[Bash: sqlite3 -readonly ~/Library/Messages/chat.db "SELECT ... WHERE text LIKE '%madhuri%' ..."]
[Bash: wacli messages search "madhuri" --limit 10]

Skip any source whose CLI isn't installed. Report which were skipped in the summary.

Presentation

Group results by source. Within each source, sort by date descending. Format:

memex: madhuri  (18 hits across 4 sources)

── logseq (9)
  (no date)  [Page] Madhuri Sakhi
  (no date)  block content snippet…
  …

── drive (3)
  2026-04-18  Madhuri + Kendre [Google Doc]  https://docs.google.com/...
  …

── gmail (3)
  2026-03-24  [sender] subject / snippet  https://mail.google.com/...
  …

── imessage (3)
  2026-03-21  [me] i think madhuri said she is only coming on the way back
  …

skipped: whatsapp (wacli not authed)

Truncate long snippets to ~240 chars. Show ref links where available (Drive URL, Gmail thread URL, logseq block UUID).

Follow-ups after a hit

When the user drills into a specific result, use the native follow-up for that source:

| Source | Drill-down | | ---------- | ------------------------------------------------------------------------- | | logseq | logseq block <uuid> --children or logseq page "<name>" --tree --resolve-embeds | | drive | gog drive download <id> --format txt --out /tmp/<name>.txt then Read | | gmail | gog gmail thread <threadId> for the full thread | | imessage | imsg chats → find chat rowid → imsg history --chat-id <id> --limit 50 around the timestamp | | whatsapp | wacli messages --chat <jid> around the timestamp |

Gotchas

1. iMessage needs Full Disk Access for the invoking terminal / Claude Code app — otherwise chat.db reads fail silently. System Settings → Privacy & Security → Full Disk Access.
2. iMessage LIKE '%…%' is a full table scan — not indexed. Expect 1–3 seconds on large histories. Narrow with a date range if possible.
3. Logseq search returns no dates. Follow up with logseq backlinks <page> --with-dates or logseq timeline <page> when dates matter.
4. imsg UUIDs vs page UUIDs vs block UUIDs. When following up, logseq blocks[].uuid → logseq block; pages[] names → logseq page; pages-content[].uuid is a page UUID, not a block.
5. Gmail query syntax leaks through. from:alice OR label:starred works for gmail but other sources will treat the whole string as a literal — expect zero hits elsewhere. Strip Gmail operators before passing to other sources if the user's query uses them.
6. Results are NOT deduplicated across sources — a Drive doc shared via Gmail legitimately shows up in both. Different surfaces, same content.
7. WhatsApp needs wacli auth one-time setup. which wacli only confirms the binary, not auth state. If wacli returns no results for a query that plausibly should hit, suggest the user run wacli doctor.
8. Logseq must be running with HTTP APIs server enabled. logseq doctor diagnoses. If not running, the logseq source silently returns nothing.

Extending to new sources

Easy additions later — each is a one-liner drop-in:

• Apple Notes → memo search "QUERY" --json
• Apple Reminders → remindctl list --search "QUERY" --json
• Calendar → gog calendar search "QUERY" --json --results-only
• Todoist → todoist list --filter "search: QUERY"
• Monarch transactions → mm transactions --search "QUERY" --json
• Past Claude sessions → via the session-logs skill
• Slack / Discord → via the respective skills
• Twitter bookmarks / Reddit saved → via xurl / reddit

When the user mentions a new source you should remember for next time, suggest adding a one-line entry to this skill's "Search commands" section.