Stop proposing to remember what the UI map already says

The map documents 447 settings, so without this every question about where a
control lives earns a proposal that costs a dismissal and teaches nothing.

Plugin/unraid/Tools/ai_chat_worker.php

@@ -774,8 +774,11 @@ if (vv_ai_mem_learn_enabled()) {
              . "MEMORY: <one short sentence>\n"
              . "Rules: one line, under 200 characters, stated as fact with no hedging. Not a "
              . "summary of your answer, not a restatement of the question, not anything already "
-             . "written above in what you know. Most exchanges warrant nothing — when in doubt, "
+             . "written above in what you know.\n"
-             . "leave the line out entirely.\n\n";
+             . "Never where a setting lives or how to reach it in the web UI — all of that is "
+             . "already documented and retrievable, so remembering it gains nothing and costs "
+             . "budget that a fact the documents cannot supply would have used.\n"
+             . "Most exchanges warrant nothing — when in doubt, leave the line out entirely.\n\n";
 }
 
 if (trim($mem['learned']) !== '') {

Plugin/unraid/include/ai_memory_learn.php

@@ -111,6 +111,28 @@ function vv_ai_mem_is_dup(string $text, ?array $known = null): bool {
 // rather than stating a fact, and it would eat the budget it is competing for.
 const VV_AI_MEM_MAX_LINE = 220;
 
+// Where a setting lives in the UI is documented, and documentation does not belong in memory —
+// memory is for what the documents cannot tell you. The UI map made this a real problem rather
+// than a theoretical one: it describes 447 settings, so without this the assistant will propose
+// "X is on the Settings tab" over and over, each one costing a dismissal.
+//
+// Gated on WHICH document the candidate is nearest to, not on how near. A plain score threshold
+// was tried first and does not separate these: a genuine fact about rsync scores 0.78 against
+// rsync.sh because it is about rsync, higher than the junk proposal that started this at 0.75.
+// Similarity measures topic, not novelty. But the junk lands on the UI map and nothing worth
+// remembering ever has — that is a difference in kind, and kind is what the index records.
+const VV_AI_MEM_UI_NEAR = 0.62;
+
+function vv_ai_mem_is_ui_fact(string $text): bool {
+    // Retrieval needs the index and the embedder. If either is unavailable the answer is "not
+    // provably documented", which files the proposal — the operator can still dismiss it, and
+    // failing closed here would silently stop learning whenever Ollama was restarting.
+    $r = @vv_ai_retrieve($text, '', '', 1);
+    if (empty($r['ok']) || empty($r['results'])) return false;
+    $top = $r['results'][0];
+    return ($top['kind'] ?? '') === 'ui' && (float) ($top['score'] ?? 0) >= VV_AI_MEM_UI_NEAR;
+}
+
 function vv_ai_mem_propose(string $text, array $meta = []): array {
     if (!vv_ai_mem_learn_enabled()) return ['ok' => false, 'error' => 'learning disabled'];
 
@@ -118,6 +140,7 @@ function vv_ai_mem_propose(string $text, array $meta = []): array {
     if ($text === '')                             return ['ok' => false, 'error' => 'empty'];
     if (mb_strlen($text) > VV_AI_MEM_MAX_LINE)    return ['ok' => false, 'error' => 'too long'];
     if (vv_ai_mem_is_dup($text))                  return ['ok' => false, 'error' => 'duplicate'];
+    if (vv_ai_mem_is_ui_fact($text))              return ['ok' => false, 'error' => 'already in the UI map'];
 
     $id  = substr(hash('sha256', $text . microtime(true)), 0, 12);
     $row = [