Optimizing OpenClaw's Quiz Response Speed

Date: March 14, 2026 · Problem: adding a new quiz question was taking over 2 minutes to respond.

The symptom

#

Lately, sending OpenClaw a “add to quiz” command left the bot sitting in a “typing” state for a long stretch — 2 to 3 minutes before it would respond — and the delay kept getting worse as the quiz grew.

Root cause

#

The data layout. Digging in, I found the quiz data was split across two files: quiz_state.json (56KB, 225 entries) — the actual practice engine, holding SRS data like totalCorrect/totalWrong/weight — and pt_learning_quiz.json (77KB, 224 entries) — a backup of the original question bank with detailed explanations, categories, and tags. Practice sessions only ever read quiz_state.json; the other file sat essentially unused.

Why it was slow. quiz_state.json was stored as a standard JSON array. Every time a question got added, the agent had to read the entire file (56KB, 15,000+ tokens), parse the whole array, append the new entry, and write the whole thing back. The bigger the file got, the more tokens each operation burned, and the slower it got.

Extra overhead from duplicate checking. The existing rules required checking for a duplicate ID before adding anything new, which meant scanning every entry — adding even more time to each request.

Fixes

#

Turning off duplicate checking. Added a rule to memory/quiz_rules.md:

Bulk Addition: Skip duplicate checking and verification during additions.
Simply append new items to the records.

Result: additions skip the full scan now, but still require a full read/write of the file.

Migrating to NDJSON (the core fix). Converted quiz_state.json from a standard JSON array to NDJSON — one independent JSON object per line:

{"id":"verb_chegar...","question":"...","srs":{...}}
{"id":"sentence_ele...","question":"...","srs":{...}}
{"id":"sentence_eu...","question":"...","srs":{...}}

The original file was backed up as quiz_state.json.bak (56KB); the migrated file came out at 44KB — 12KB smaller, with all 225 entries intact.

Added rules 8 and 9 to quiz_rules.md to document the new format:

8. Storage Format: NDJSON, one independent JSON object per line
9. Operations:
   - Add: append a new line directly (no need to read the existing file)
   - Update: read the full file → edit the matching line → write it back

Before and after

#

What’s left unoptimized

#

Updating SRS after answering still can’t be sped up — every answer needs to update a specific entry’s totalCorrect/totalWrong, and NDJSON still requires a full read/write since there’s no way to edit a single line in place. With 225 entries it’s not a big deal right now; if it grows into the thousands, migrating to SQLite is worth considering.

pt_learning_quiz.json stays as a backup — I’m not folding its explanation field into quiz_state.json, to avoid bloating that file further.

Files involved

#

/root/clawd/memory/quiz_state.json        # NDJSON format, currently in use
/root/clawd/memory/quiz_state.json.bak    # backup of the original JSON array format
/root/clawd/memory/pt_learning_quiz.json  # original question bank backup (unused)
/root/clawd/memory/quiz_rules.md          # SRS rules file (updated)