diff --git a/core/cognition/vram_orchestrator.py b/core/cognition/vram_orchestrator.py
new file mode 100644
index 000000000..5e5d5962c
--- /dev/null
+++ b/core/cognition/vram_orchestrator.py
@@ -0,0 +1,191 @@
+"""
+Orchestrateur VRAM — gère le chargement/déchargement des modèles selon le mode.
+
+Deux modes :
+- SHADOW : streaming server + agent_chat actifs, VLM raisonnement déchargé
+- REPLAY : VLM raisonnement (qwen2.5vl:7b) chargé, services non-essentiels stoppés
+
+Bascule automatique ou manuelle selon le contexte.
+"""
+
+import logging
+import os
+import subprocess
+import time
+from enum import Enum
+from typing import Optional
+
+logger = logging.getLogger(__name__)
+
+OLLAMA_URL = os.environ.get("OLLAMA_URL", "http://localhost:11434")
+REASONING_MODEL = os.environ.get("RPA_REASONING_MODEL", "qwen2.5vl:7b")
+MIN_VRAM_FOR_REASONING = 5.0  # Go minimum pour charger le modèle de raisonnement
+
+
+class VRAMMode(Enum):
+    SHADOW = "shadow"
+    REPLAY = "replay"
+
+
+class VRAMOrchestrator:
+    """Gère la VRAM pour éviter les conflits entre modèles."""
+
+    def __init__(self):
+        self._current_mode: Optional[VRAMMode] = None
+        self._stopped_services: list = []
+
+    def get_free_vram_gb(self) -> float:
+        """Retourne la VRAM libre en Go."""
+        try:
+            result = subprocess.run(
+                ["nvidia-smi", "--query-gpu=memory.free", "--format=csv,noheader,nounits"],
+                capture_output=True, text=True, timeout=5
+            )
+            return float(result.stdout.strip()) / 1024
+        except Exception:
+            return 0.0
+
+    def get_used_vram_gb(self) -> float:
+        """Retourne la VRAM utilisée en Go."""
+        try:
+            result = subprocess.run(
+                ["nvidia-smi", "--query-gpu=memory.used", "--format=csv,noheader,nounits"],
+                capture_output=True, text=True, timeout=5
+            )
+            return float(result.stdout.strip()) / 1024
+        except Exception:
+            return 0.0
+
+    def switch_to_replay(self) -> bool:
+        """Bascule en mode replay : libère la VRAM pour le VLM de raisonnement.
+
+        1. Stoppe les services non-essentiels (agent_chat)
+        2. Redémarre Ollama pour libérer les modèles chargés
+        3. Précharge le modèle de raisonnement
+        """
+        if self._current_mode == VRAMMode.REPLAY:
+            logger.info("Déjà en mode REPLAY")
+            return True
+
+        logger.info("Bascule en mode REPLAY...")
+
+        # Stopper agent_chat si il tourne
+        try:
+            result = subprocess.run(
+                ["pgrep", "-f", "agent_chat"],
+                capture_output=True, text=True, timeout=5
+            )
+            pids = result.stdout.strip().split('\n')
+            for pid in pids:
+                if pid.strip():
+                    subprocess.run(["kill", pid.strip()], timeout=5)
+                    self._stopped_services.append(("agent_chat", pid.strip()))
+                    logger.info(f"agent_chat stoppé (PID {pid.strip()})")
+        except Exception as e:
+            logger.debug(f"Pas d'agent_chat à stopper: {e}")
+
+        # Redémarrer Ollama pour libérer la mémoire
+        try:
+            subprocess.run(["sudo", "systemctl", "restart", "ollama"],
+                         timeout=10, check=True)
+            time.sleep(2)
+            logger.info("Ollama redémarré")
+        except Exception as e:
+            logger.warning(f"Impossible de redémarrer Ollama: {e}")
+
+        # Vérifier la VRAM disponible
+        free = self.get_free_vram_gb()
+        logger.info(f"VRAM libre: {free:.1f} Go")
+
+        if free < MIN_VRAM_FOR_REASONING:
+            logger.warning(f"VRAM insuffisante ({free:.1f} Go < {MIN_VRAM_FOR_REASONING} Go)")
+            return False
+
+        # Précharger le modèle de raisonnement
+        try:
+            import requests
+            logger.info(f"Préchargement {REASONING_MODEL}...")
+            resp = requests.post(f"{OLLAMA_URL}/api/generate", json={
+                "model": REASONING_MODEL,
+                "prompt": "test",
+                "stream": False,
+                "options": {"num_predict": 1}
+            }, timeout=60)
+            if resp.status_code == 200:
+                logger.info(f"{REASONING_MODEL} chargé en VRAM")
+            free_after = self.get_free_vram_gb()
+            logger.info(f"VRAM libre après chargement: {free_after:.1f} Go")
+        except Exception as e:
+            logger.warning(f"Préchargement échoué: {e}")
+
+        self._current_mode = VRAMMode.REPLAY
+        return True
+
+    def switch_to_shadow(self) -> bool:
+        """Bascule en mode shadow : relance les services d'observation.
+
+        1. Redémarre Ollama (décharge le VLM de raisonnement)
+        2. Relance les services stoppés
+        """
+        if self._current_mode == VRAMMode.SHADOW:
+            logger.info("Déjà en mode SHADOW")
+            return True
+
+        logger.info("Bascule en mode SHADOW...")
+
+        # Redémarrer Ollama
+        try:
+            subprocess.run(["sudo", "systemctl", "restart", "ollama"],
+                         timeout=10, check=True)
+            time.sleep(2)
+        except Exception as e:
+            logger.warning(f"Impossible de redémarrer Ollama: {e}")
+
+        # Relancer les services stoppés
+        for service_name, _pid in self._stopped_services:
+            try:
+                if service_name == "agent_chat":
+                    subprocess.Popen(
+                        ["python3", "-m", "agent_chat.app"],
+                        cwd="/home/dom/ai/rpa_vision_v3",
+                        stdout=subprocess.DEVNULL,
+                        stderr=subprocess.DEVNULL
+                    )
+                    logger.info(f"{service_name} relancé")
+            except Exception as e:
+                logger.warning(f"Impossible de relancer {service_name}: {e}")
+
+        self._stopped_services.clear()
+        self._current_mode = VRAMMode.SHADOW
+        return True
+
+    def ensure_reasoning_ready(self) -> bool:
+        """Vérifie que le VLM de raisonnement est prêt. Bascule si nécessaire."""
+        free = self.get_free_vram_gb()
+        if free >= MIN_VRAM_FOR_REASONING:
+            return True
+        return self.switch_to_replay()
+
+    @property
+    def current_mode(self) -> Optional[str]:
+        return self._current_mode.value if self._current_mode else None
+
+    def status(self) -> dict:
+        return {
+            "mode": self.current_mode,
+            "vram_free_gb": round(self.get_free_vram_gb(), 1),
+            "vram_used_gb": round(self.get_used_vram_gb(), 1),
+            "reasoning_model": REASONING_MODEL,
+            "stopped_services": [s[0] for s in self._stopped_services],
+        }
+
+
+# Singleton
+_orchestrator: Optional[VRAMOrchestrator] = None
+
+
+def get_orchestrator() -> VRAMOrchestrator:
+    global _orchestrator
+    if _orchestrator is None:
+        _orchestrator = VRAMOrchestrator()
+    return _orchestrator
diff --git a/core/execution/input_handler.py b/core/execution/input_handler.py
index 82cc8b375..bc0816295 100644
--- a/core/execution/input_handler.py
+++ b/core/execution/input_handler.py
@@ -286,8 +286,12 @@ Si tu vois un dialogue ou une popup, indique quel bouton cliquer.
 Si l'écran est normal sans action nécessaire, réponds action="nothing".
 Réponds UNIQUEMENT le JSON, pas d'explication."""
 
+        from core.cognition.vram_orchestrator import get_orchestrator
+        orch = get_orchestrator()
+        orch.ensure_reasoning_ready()
+
         ollama_url = os.environ.get("OLLAMA_URL", "http://localhost:11434")
-        model = os.environ.get("RPA_REASONING_MODEL", os.environ.get("RPA_VLM_MODEL", "qwen2.5vl:3b"))
+        model = os.environ.get("RPA_REASONING_MODEL", "qwen2.5vl:7b")
 
         response = requests.post(
             f"{ollama_url}/api/generate",