memotag-rs

Rust daemon for Memos: content-based auto-tagging plus CalDAV task/event sync. Prod runs on ubuntu/oracle as ghcr.io/0124212/memotag-rs:main next to memos (neosmemo/memos:stable).

Tag rules (content → #hashtag appended to the memo)

Memos 0.30 treats tags as output-only (parsed from #tags in content), so the tagger PATCHes content, never a tags field.

• code — any fenced block (```, bare or any language)
• diagram — fenced ```mermaid blocks
• math — $$…$$ or \frac-style commands
• event — 📅 YYYY-MM-DD lines (same parser as the CalDAV sync)
• task — - [ ] / - [x] anywhere: indented, */+, numbered
• quote — > blockquote on any line, incl. indented
• link — bare https://… URLs; known video hosts also imply video
• image — markdown ![](…), <img>, image extensions (also via attachments)
• figure — markdown-image syntax or <figure>/<img> HTML
• table — |…| rows plus a --- separator
• audio / video — media extensions (also via attachments)
• <ext> — known file extensions (pdf, rs, mp3, …) as extra tags
• fallback inbox (configurable) only for memos with no tags at all

Folds (single write per memo): #untagged→#inbox, #tasks→#task, legacy plurals (#links, #images, …)→singular; strips stale #inbox once real tags exist. Curated memos (real tags, no inbox) are never force-tagged with the fallback, but new rule hits still backfill.

Known server quirk: a doc containing HTML (<figure>, <img>…) gets tags=null back — the 0.30 extractor bails on the whole doc. Hashtags stay in content (searchable) and converge (no rewrite loop).

Run

export MEMOS_URL="http://memos:5230"          # prod (from inside compose net)
export MEMOS_API_TOKEN="memos_pat_xxx"        # required
export AUTOTAG_DEFAULT_TAG="inbox"            # optional
export AUTOTAG_INTERVAL=300                   # optional, seconds
export AUTOTAG_CONCURRENCY=8                  # optional, 1-32 (AUTOTAG_BATCH_SIZE also read)

cargo run --release -- --once --dry-run  # test pass: log only, no writes
cargo run --release -- --once             # single live pass, then exit
cargo run --release -- --clean --dry-run  # preview junk-hashtag cleanup
memotag-rs                                # daemon: interval loop + webhook server

Testing (debian dummy, before touching ubuntu)

Prod lives on ubuntu — never test against it. Use the AI-controlled dummy:

cd /root/memos-test
docker compose up -d     # http://127.0.0.1:5230, same image as prod
python3 seed.py          # 27 edge-case memos + typo/good fixtures
MEMOS_URL=http://127.0.0.1:5230 MEMOS_API_TOKEN=$MEMOS_TEST_PAT \
  /root/code/memotag-rs/target/release/memotag-rs --once --dry-run
MEMOS_URL=http://127.0.0.1:5230 MEMOS_API_TOKEN=$MEMOS_TEST_PAT \
  /root/code/memotag-rs/target/release/memotag-rs --once
python3 verify.py        # must PASS before any ubuntu deploy

cargo test runs 18 autotag unit tests (no server needed). Two unrelated tests (caldav vevent, parser priority) fail on main — pre-existing, untouched.

Pitfall hunting (scan binary, not shipped in the image)

cargo run --bin scan -- /root/memos-test/prod-copy/memos_prod_copy.db

Runs the tagger's pure logic over a read-only DB copy: would-change counts, per-tag backfill stats, URL-fragment phantoms, fold-targets-inside-fences, and — critically — non-convergent memos (output that would rewrite forever). Exit 1 on non-convergence or fence hazards. Proven on prod data: 40 would-change / 0 non-convergent / 0 fence hazards.

Deploy (ubuntu)

Push main → GHCR builds :main → on oracle: docker compose pull memotag-rs && docker compose up -d memotag-rs.