Improve pgvector keyword recall ranking without changing hybrid sort.

Rank ILIKE matches by lexical coverage instead of recency and attach lexical scores to keyword rows so MemFuseBench candidate recall rises sharply while Chinese hybrid ordering stays unchanged.

Co-authored-by: Cursor <cursoragent@cursor.com>

memory-v2-pgvector.mjs

@@ -85,6 +85,8 @@ function dedupeContentPrefix(text, length = 96) {
   return normalizeSearchText(String(text ?? '').slice(0, length));
 }
 
+const KEYWORD_FETCH_CAP = 500;
+
 async function fetchKeywordCandidates(pool, {
   userId,
   query,
@@ -96,17 +98,32 @@ async function fetchKeywordCandidates(pool, {
   if (!terms.length) return [];
   const clauses = terms.map((_term, index) => `content ILIKE $${index + 2}`);
   const params = [userId, ...terms.map((term) => `%${term}%`)];
+  const safeLimit = Math.max(1, Math.min(50, Number(limit) || 20));
+  const fetchCap = Math.min(KEYWORD_FETCH_CAP, Math.max(safeLimit, safeLimit * 10));
   const sql = `
     SELECT id, content, type, created_at, updated_at, 1.0 AS score
     FROM ${tableName}
     WHERE user_id = $1
       AND (${clauses.join(' OR ')})
-    ORDER BY updated_at DESC
     LIMIT $${params.length + 1}
   `;
-  params.push(Math.max(1, Math.min(50, Number(limit) || 20)));
+  params.push(fetchCap);
   const result = await pool.query(sql, params);
-  return result?.rows ?? [];
+  const rows = result?.rows ?? [];
+  return rows
+    .map((row) => ({
+      row,
+      lexicalScore: lexicalQueryCoverage(query, row.content ?? ''),
+      updatedAt: timestampValue(row.updated_at ?? row.created_at),
+    }))
+    .sort((left, right) => {
+      if (left.lexicalScore !== right.lexicalScore) {
+        return right.lexicalScore - left.lexicalScore;
+      }
+      return right.updatedAt - left.updatedAt;
+    })
+    .slice(0, safeLimit)
+    .map(({ row, lexicalScore }) => ({ ...row, score: lexicalScore }));
 }
 
 function timestampValue(value) {