Public AI moments are the observation only: the raw speech-to-text clause (" — heard: …", up to 500 characters, often another language) stays for the AI's context and is never shown (home 'Right now', a VOD's AI timeline, a streamer's AI profile); pose and expression wording goes like looks (owner, 2026-10-17: a Greek stream's transcript filled the home page) (#236)

ddiff --git a/server/ai/chat-ai-routes.js b/server/ai/chat-ai-routes.js
iindex c3048333..18178043 100644
---- a/server/ai/chat-ai-routes.js
++++ b/server/ai/chat-ai-routes.js
@@@ -47,7 +47,7 @@ router.get('/user/:id', async (req, res) => {
                  // for more, and default high enough to read as a real timeline.
                  const memLimit = Math.min(200, Math.max(1, parseInt(req.query.memories, 10) || 40));
                  const mems = (await db.getStreamMemoriesByUser(uid, memLimit) || []).map(m => ({
--                    description: m.description || '',
++                    description: db.publicMemoryText(m.description),
                      created_at: m.created_at,
                      stream_id: m.stream_id,
                      offset_seconds: m.offset_seconds,
ddiff --git a/server/db/database.js b/server/db/database.js
iindex 63c5810d..866abf64 100644
---- a/server/db/database.js
++++ b/server/db/database.js
@@@ -433,9 +433,11 @@ const _hasLooks = (t) => _LOOKS_WORD.test(String(t || ''));
  function _noLooksTags(tags) {
      return (Array.isArray(tags) ? tags : []).filter((t) => typeof t === 'string' && t.trim() && !_LOOKS_TAG.test(t.trim()));
  }
++// Body language is the person too ("reacting with a thoughtful pose", "in a relaxed posture"): the phrase goes.
++const _POSE = /,?\s*(?:(?:appearing|seeming|seemingly)\s+(?:to\s+be\s+)?)?(?:(?:reacting|sitting|standing|leaning|posing|looking|resting)\s+)?(?:with|in)\s+(?:an?|their|his|her)\s+(?:[\w-]+\s+){0,2}(?:pose|posture|expression|facial expression|stance)\b/gi;
  function _noLooks(text) {
      if (!text || typeof text !== 'string') return text;
--    return _noClothing(text)
++    return _noClothing(text).replace(_POSE, '').replace(/\s+([,.;:!?])/g, '$1').replace(/ {2,}/g, ' ')
          // Looks named: the person on camera. Only an age ("a child sits nearby"): someone, who may not be the streamer.
          .replace(_LOOKED_AT, (m, art, looks, noun) => (looks ? _cap(m, 'the streamer') : _AGE_NOUN.test(noun) ? _cap(m, /^the$/i.test(art) ? 'the person' : 'someone') : m))
          .replace(_LOOKED_AT_BARE, (m) => _cap(m, 'streamer'))
@@@ -446,6 +448,16 @@ function _noLooks(text) {
          .replace(/\s*,\s*(?:shirtless|topless|bare-chested)\s*,/gi, '')
          .replace(/\b(?:shirtless|topless|bare-chested)\b,?\s*/gi, '');
  }
++/**
++ * A stream memory as the public sees it (home "Right now", a VOD's AI timeline, a streamer's AI profile): the
++ * observation only. The " — heard: "…"" clause is raw speech-to-text kept for the AI's own context (often another
++ * language, misheard, 500 characters); shown, it read as a wall of garbled text (2026-10-17, a Greek stream on the
++ * home page). Looks and pose wording go as everywhere public.
++ */
++function _publicMemoryText(text) {
++    const t = String(text || '').replace(/\s+—\s+heard:[\s\S]*$/, '').trim();
++    return (_noLooks(t) || '').trim();
++}
  // Derive a concise short overview from a long one — the lead sentence(s), capped
  // ~150 chars at a sentence/word boundary. Deterministic + free (no AI call), so it
  // can be cached at write time and shown on listing cards.
@@@ -3586,7 +3598,7 @@ module.exports = {
      setVodAiOverview, setClipAiOverview, setVodTranscript, setClipTranscript, getStreamMemoriesInRange,
      getVodsNeedingOverview, getClipsNeedingOverview, getVodsNeedingTimeline, getVodsNeedingTranscript, getClipsNeedingTranscript, getMediaNeedingAiAnalysis,
      setVodTranscriptStatus, setClipTranscriptStatus, bumpVodTranscriptAttempt, bumpClipTranscriptAttempt,
--    cleanupMalformedAiText, noLooks: _noLooks, noLooksTags: _noLooksTags, hasLooks: _hasLooks, recordAiUsage, getAiCostToday, getAiCostTodayForUser, getAiUsageSummary,
++    cleanupMalformedAiText, noLooks: _noLooks, noLooksTags: _noLooksTags, hasLooks: _hasLooks, publicMemoryText: _publicMemoryText, recordAiUsage, getAiCostToday, getAiCostTodayForUser, getAiUsageSummary,
      getStreamMemoriesByUser, countStreamMemoriesByUser, getStreamTranscriptSegments,
      addTimelineEvents, getTimeline, getTimelineText, getTimelineCoverage, linkTimelineToVod, getTimelineByVod, getTimelineVodId,
      upsertStreamerOverview, getStreamerOverview, getAllStreamerOverviews, getStreamersNeedingOverview,
ddiff --git a/server/home/routes.js b/server/home/routes.js
iindex 04bd73f5..29103fa4 100644
---- a/server/home/routes.js
++++ b/server/home/routes.js
@@@ -273,7 +273,8 @@ router.get('/featured', async (req, res) => {
          let moment = null;
          try {
              const m = await db.getLatestStreamMemory(pick.id);
--            if (m && m.description) moment = { description: m.description, captured_at: m.captured_at || m.created_at || null, thumbnail_url: m.thumbnail_url || null };
++            const text = m && db.publicMemoryText(m.description);
++            if (text) moment = { description: text, captured_at: m.captured_at || m.created_at || null, thumbnail_url: m.thumbnail_url || null };
          } catch { /* */ }
          res.json({ stream, moment, count: live.length, index: idx, next_in_ms: FEATURED_ROTATE_MS - (Date.now() % FEATURED_ROTATE_MS) });
      } catch (err) {
ddiff --git a/server/media-proxy/vods.js b/server/media-proxy/vods.js
iindex b8fd5428..0befc664 100644
---- a/server/media-proxy/vods.js
++++ b/server/media-proxy/vods.js
@@@ -390,7 +390,7 @@ router.get('/:id/memories', optionalAuth, async (req, res) => {
          if (!streamId) return res.json({ memories: [] });
          const memories = (await db.getStreamMemories(streamId) || []).map(m => ({
              offset_seconds: m.offset_seconds,
--            description: m.description,
++            description: db.publicMemoryText(m.description),
              tags: m.tags ? (() => { try { return JSON.parse(m.tags); } catch { return []; } })() : [],
          }));
          res.json({ memories });
ddiff --git a/test/ai-text-no-looks.test.js b/test/ai-text-no-looks.test.js
iindex d742ebf4..9e8ab6c0 100644
---- a/test/ai-text-no-looks.test.js
++++ b/test/ai-text-no-looks.test.js
@@@ -27,6 +27,10 @@ const db = require('../server/db/database');
          ['The view is close to the camera while wearing a black beanie and hoodie, with trees around.', 'The view is close to the camera, with trees around.'],
          ['A man in a gray tank top sits against a plain light background.', 'A man sits against a plain light background.'],
          ['A man with a long beard and plaid shirt examines a woven item.', 'A man examines a woven item.'],
++        // Pose and expression are the person too (2026-10-17, the home page's "Right now").
++        ['A live stream frame shows the streamer in a small window bottom-left, appearing to be reacting with a thoughtful pose while the main screen is black.',
++            'A live stream frame shows the streamer in a small window bottom-left while the main screen is black.'],
++        ['The streamer sits in a relaxed posture at the desk.', 'The streamer sits at the desk.'],
          ['A kitchen studio with a headset-wearing host, microphone, and on-screen graphics.', 'A kitchen studio with a host, microphone, and on-screen graphics.'],
          ['Panelists discuss while wearing headsets and microphones, with a trophy nearby.', 'Panelists discuss, with a trophy nearby.'],
          ['A man wearing sunglasses on his head looks down at the camera.', 'A man looks down at the camera.'],
@@@ -46,6 +50,9 @@ const db = require('../server/db/database');
          ['The streamer laughs — heard: "Put your headset on. I like your hair."', 'The streamer laughs — heard: "Put your headset on. I like your hair."'],
      ];
      let off = 0;
++    // Public memory text is the observation alone: the raw transcript clause stays for the AI's context, never shown.
++    assert.strictEqual(db.publicMemoryText('The main screen is black, suggesting a pause. — heard: "Δεν θα πούμε άλλο. Τώρα που κερδίζουμε."'), 'The main screen is black, suggesting a pause.');
++    assert.strictEqual(db.publicMemoryText(null), '');
      for (const [input, want] of cases) {
          await db.addStreamMemory({ stream_id: sid, user_id: uid, offset_seconds: (off += 60), description: input });
          const row = await db.get('SELECT description FROM stream_memories WHERE stream_id = ? AND offset_seconds = ?', [sid, off]);