diff --git a/skills/cognitive-memory/client.py b/skills/cognitive-memory/client.py
index 8d0b513..b563cf9 100644
--- a/skills/cognitive-memory/client.py
+++ b/skills/cognitive-memory/client.py
@@ -410,14 +410,16 @@ def calculate_decay_score(
 
 
 def _ollama_embed(
-    texts: List[str], timeout: int = EMBEDDING_TIMEOUT
+    texts: List[str],
+    model: str = EMBEDDING_MODEL,
+    timeout: int = EMBEDDING_TIMEOUT,
 ) -> Optional[List[List[float]]]:
     """Get embeddings from Ollama for a list of texts.
 
     Returns list of embedding vectors, or None if Ollama is unavailable.
     """
     try:
-        payload = json.dumps({"model": EMBEDDING_MODEL, "input": texts}).encode("utf-8")
+        payload = json.dumps({"model": model, "input": texts}).encode("utf-8")
         req = urllib.request.Request(
             f"{OLLAMA_URL}/api/embed",
             data=payload,
@@ -2050,13 +2052,13 @@ class CognitiveMemoryClient:
                     return vectors, "openai", model
             # Fallback to ollama
             ollama_model = config.get("ollama_model", EMBEDDING_MODEL)
-            vectors = _ollama_embed(texts, timeout=timeout)
+            vectors = _ollama_embed(texts, model=ollama_model, timeout=timeout)
             if vectors is not None:
                 return vectors, "ollama", ollama_model
         else:
             # ollama first
             ollama_model = config.get("ollama_model", EMBEDDING_MODEL)
-            vectors = _ollama_embed(texts, timeout=timeout)
+            vectors = _ollama_embed(texts, model=ollama_model, timeout=timeout)
             if vectors is not None:
                 return vectors, "ollama", ollama_model
             # Fallback to openai
@@ -2069,11 +2071,15 @@ class CognitiveMemoryClient:
 
         return None, "", ""
 
-    def embed(self) -> Dict[str, Any]:
+    def embed(self, if_changed: bool = False) -> Dict[str, Any]:
         """Generate embeddings for all memories using configured provider.
 
         Detects provider changes and re-embeds everything (dimension mismatch safety).
         Stores vectors in _embeddings.json (not git-tracked).
+
+        Args:
+            if_changed: If True, skip embedding if the set of memory IDs hasn't
+                changed since last run (no new/deleted memories).
         """
         index = self._load_index()
         entries = index.get("entries", {})
@@ -2097,12 +2103,29 @@ class CognitiveMemoryClient:
 
         config = self._get_embedding_provider()
         new_provider = config.get("embedding_provider", "ollama")
-        if old_provider and old_provider != new_provider:
+        provider_changed = old_provider and old_provider != new_provider
+        if provider_changed:
             print(
                 f"Provider changed ({old_provider} -> {new_provider}), re-embedding all memories...",
                 file=sys.stderr,
             )
 
+        # Skip if nothing changed (unless provider switched)
+        if if_changed and not provider_changed and embeddings_path.exists():
+            try:
+                old_data = json.loads(embeddings_path.read_text())
+                embedded_ids = set(old_data.get("entries", {}).keys())
+                index_ids = set(entries.keys())
+                if embedded_ids == index_ids:
+                    return {
+                        "embedded": 0,
+                        "skipped": True,
+                        "reason": "no new or deleted memories",
+                        "path": str(embeddings_path),
+                    }
+            except (json.JSONDecodeError, OSError):
+                pass  # Can't read old data, re-embed
+
         # Build texts to embed
         memory_ids = list(entries.keys())
         texts = []
@@ -2180,7 +2203,8 @@ class CognitiveMemoryClient:
                 if vecs:
                     query_vec = vecs[0]
         if query_vec is None and stored_provider == "ollama":
-            vecs = _ollama_embed([query])
+            stored_model = emb_data.get("model", EMBEDDING_MODEL)
+            vecs = _ollama_embed([query], model=stored_model)
             if vecs:
                 query_vec = vecs[0]
         # Last resort: try any available provider
@@ -2896,9 +2920,14 @@ def main():
     subparsers.add_parser("reindex", help="Rebuild index from files")
 
     # embed
-    subparsers.add_parser(
+    embed_parser = subparsers.add_parser(
         "embed", help="Generate embeddings for all memories via Ollama"
     )
+    embed_parser.add_argument(
+        "--if-changed",
+        action="store_true",
+        help="Skip if no memories were added or deleted since last embed",
+    )
 
     # pin
     sp = subparsers.add_parser("pin", help="Move memory to vault (never decays)")
@@ -2958,6 +2987,9 @@ def main():
         "--provider", choices=["ollama", "openai"], help="Set embedding provider"
     )
     sp.add_argument("--openai-key", help="Set OpenAI API key")
+    sp.add_argument(
+        "--ollama-model", help="Set Ollama model name (e.g. qwen3-embedding:8b)"
+    )
 
     args = parser.parse_args()
 
@@ -3119,11 +3151,13 @@ def main():
         result = {"success": True, "indexed": count}
 
     elif args.command == "embed":
-        print(
-            "Generating embeddings (this may take a while if model needs to be pulled)...",
-            file=sys.stderr,
-        )
-        result = client.embed()
+        if_changed = getattr(args, "if_changed", False)
+        if not if_changed:
+            print(
+                "Generating embeddings (this may take a while if model needs to be pulled)...",
+                file=sys.stderr,
+            )
+        result = client.embed(if_changed=if_changed)
 
     elif args.command == "pin":
         success = client.pin(args.memory_id)
@@ -3204,6 +3238,9 @@ def main():
         if args.openai_key:
             config["openai_api_key"] = args.openai_key
             changed = True
+        if args.ollama_model:
+            config["ollama_model"] = args.ollama_model
+            changed = True
 
         if changed:
             config_path.write_text(json.dumps(config, indent=2))