diff --git a/stt/server/pyproject.toml b/stt/server/pyproject.toml
index e992fb3..86202f7 100644
--- a/stt/server/pyproject.toml
+++ b/stt/server/pyproject.toml
@@ -1,6 +1,6 @@
 [project]
 name = "stt-server"
-version = "0.1.0"
+version = "0.2.0"
 description = "STT-server — orchestrateur AI du homelab Funk (API pour les clients STT)"
 requires-python = ">=3.11"
 
diff --git a/stt/server/stt_server/__init__.py b/stt/server/stt_server/__init__.py
index f614a39..fe833f3 100644
--- a/stt/server/stt_server/__init__.py
+++ b/stt/server/stt_server/__init__.py
@@ -1,3 +1,3 @@
 """STT-server — orchestrateur AI in-cluster pour les clients STT."""
 
-__version__ = "0.1.0"
+__version__ = "0.2.0"
diff --git a/stt/server/stt_server/app.py b/stt/server/stt_server/app.py
index 9865e72..f3fa7e9 100644
--- a/stt/server/stt_server/app.py
+++ b/stt/server/stt_server/app.py
@@ -7,21 +7,39 @@ Endpoints :
 
 from __future__ import annotations
 
+import logging
+import time
+from contextlib import asynccontextmanager
+
 import httpx
-from fastapi import FastAPI, HTTPException
+from fastapi import BackgroundTasks, FastAPI, HTTPException
 from pydantic import BaseModel
 
 from stt_server import __version__
+from stt_server import brain
 from stt_server.brain import ask as brain_ask
 from stt_server.config import settings
 from stt_server.longterm import LongTermMemory
 from stt_server.memory import SessionStore
 
-app = FastAPI(title="STT-server", version=__version__)
+log = logging.getLogger("stt_server")
+
 sessions = SessionStore()
 longterm = LongTermMemory() if settings.memory_longterm else None
 
 
+@asynccontextmanager
+async def lifespan(app: FastAPI):
+    yield
+    # Fermeture propre des clients HTTP persistants (pooling).
+    await brain.aclose()
+    if longterm:
+        await longterm.aclose()
+
+
+app = FastAPI(title="STT-server", version=__version__, lifespan=lifespan)
+
+
 class AskRequest(BaseModel):
     text: str
     model: str | None = None      # alias LiteLLM ; défaut serveur si absent
@@ -59,7 +77,7 @@ async def v1_reset(req: AskRequest) -> dict:
 
 
 @app.post("/v1/ask", response_model=AskReply)
-async def v1_ask(req: AskRequest) -> AskReply:
+async def v1_ask(req: AskRequest, background: BackgroundTasks) -> AskReply:
     text = req.text.strip()
     if not text:
         raise HTTPException(status_code=400, detail="text vide")
@@ -69,17 +87,30 @@ async def v1_ask(req: AskRequest) -> AskReply:
             status_code=400,
             detail=f"modèle '{model}' non autorisé ; dispo : {settings.allowed_models}",
         )
+    t0 = time.perf_counter()
     history = sessions.history(req.session_id) if req.session_id else None
-    memories = await longterm.recall(text) if longterm else None
+    # recall : timeout serré, dégrade vite ; renvoie aussi le vecteur (réutilisé par store)
+    memories, qvec = await longterm.recall(text) if longterm else ([], None)
+    t_recall = time.perf_counter()
     try:
         reply = await brain_ask(text, model, history, memories)
     except httpx.HTTPError as e:
         raise HTTPException(status_code=502, detail=f"upstream LiteLLM : {e}") from e
+    t_gen = time.perf_counter()
     if req.session_id:
         sessions.add(req.session_id, "user", text)
         sessions.add(req.session_id, "assistant", reply)
+    # store : APRÈS la réponse (BackgroundTasks) → hors latence perçue, et on réutilise qvec
     if longterm:
-        await longterm.remember(req.session_id or "anon", text)
+        background.add_task(longterm.store, req.session_id or "anon", text, qvec)
+    log.info(
+        "ask model=%s recall=%.0fms gen=%.0fms total=%.0fms mem=%d",
+        model,
+        (t_recall - t0) * 1000,
+        (t_gen - t_recall) * 1000,
+        (t_gen - t0) * 1000,
+        len(memories),
+    )
     return AskReply(reply=reply, model=model)
 
 
diff --git a/stt/server/stt_server/brain.py b/stt/server/stt_server/brain.py
index 098dbf4..8fdfe6f 100644
--- a/stt/server/stt_server/brain.py
+++ b/stt/server/stt_server/brain.py
@@ -11,6 +11,23 @@ import httpx
 
 from stt_server.config import settings
 
+# Client persistant (pooling + keep-alive) : évite un handshake TCP vers LiteLLM à chaque tour.
+_client: httpx.AsyncClient | None = None
+
+
+def _get_client() -> httpx.AsyncClient:
+    global _client
+    if _client is None or _client.is_closed:
+        _client = httpx.AsyncClient(timeout=settings.request_timeout)
+    return _client
+
+
+async def aclose() -> None:
+    global _client
+    if _client is not None and not _client.is_closed:
+        await _client.aclose()
+    _client = None
+
 
 async def ask(
     text: str,
@@ -42,13 +59,12 @@ async def ask(
         "temperature": settings.temperature,
     }
     headers = {"Authorization": f"Bearer {settings.litellm_key}"}
-    async with httpx.AsyncClient(timeout=settings.request_timeout) as client:
-        r = await client.post(settings.litellm_url, json=payload, headers=headers)
-        r.raise_for_status()
-        msg = r.json()["choices"][0]["message"]
-        # Filet de sécurité : si un modèle « thinking » renvoie un content vide (tout parti
-        # en reasoning_content), on récupère le raisonnement plutôt que de renvoyer "".
-        content = (msg.get("content") or "").strip()
-        if not content:
-            content = (msg.get("reasoning_content") or "").strip()
-        return content
+    r = await _get_client().post(settings.litellm_url, json=payload, headers=headers)
+    r.raise_for_status()
+    msg = r.json()["choices"][0]["message"]
+    # Filet de sécurité : si un modèle « thinking » renvoie un content vide (tout parti
+    # en reasoning_content), on récupère le raisonnement plutôt que de renvoyer "".
+    content = (msg.get("content") or "").strip()
+    if not content:
+        content = (msg.get("reasoning_content") or "").strip()
+    return content
diff --git a/stt/server/stt_server/config.py b/stt/server/stt_server/config.py
index c1c7778..e6ed3b6 100644
--- a/stt/server/stt_server/config.py
+++ b/stt/server/stt_server/config.py
@@ -43,6 +43,11 @@ class Settings:
     embed_url: str = os.getenv("STT_EMBED_URL", "http://192.168.10.20:1238/v1/embeddings")
     embed_model: str = os.getenv("STT_EMBED_MODEL", "nomic-embed-text")
     memory_top_k: int = int(os.getenv("STT_MEMORY_TOPK", "3"))
+    # Le recall (embed + recherche) est sur le chemin de réponse : timeout SERRÉ pour qu'un
+    # embed lent/mort dégrade vite (souvenirs vides) au lieu d'ajouter des secondes au client.
+    # Le store tourne en tâche de fond (après la réponse) → timeout plus large toléré.
+    memory_recall_timeout: float = float(os.getenv("STT_MEMORY_RECALL_TIMEOUT", "4"))
+    memory_store_timeout: float = float(os.getenv("STT_MEMORY_STORE_TIMEOUT", "20"))
 
 
 settings = Settings()
diff --git a/stt/server/stt_server/longterm.py b/stt/server/stt_server/longterm.py
index 93d65e1..84a38b1 100644
--- a/stt/server/stt_server/longterm.py
+++ b/stt/server/stt_server/longterm.py
@@ -27,21 +27,37 @@ class LongTermMemory:
         self.embed_url = settings.embed_url
         self.embed_model = settings.embed_model
         self.top_k = settings.memory_top_k
+        self.recall_timeout = settings.memory_recall_timeout
+        self.store_timeout = settings.memory_store_timeout
         self._ready = False
+        # Client persistant : pooling + keep-alive (évite un handshake TCP par appel).
+        self._http: httpx.AsyncClient | None = None
 
-    async def _embed(self, client: httpx.AsyncClient, text: str) -> list[float]:
-        r = await client.post(
+    def _client(self) -> httpx.AsyncClient:
+        if self._http is None or self._http.is_closed:
+            self._http = httpx.AsyncClient()
+        return self._http
+
+    async def aclose(self) -> None:
+        if self._http is not None and not self._http.is_closed:
+            await self._http.aclose()
+
+    async def _embed(self, text: str, timeout: float) -> list[float]:
+        r = await self._client().post(
             self.embed_url,
             json={"model": self.embed_model, "input": text},
-            timeout=30,
+            timeout=timeout,
         )
         r.raise_for_status()
         return r.json()["data"][0]["embedding"]
 
-    async def _ensure_collection(self, client: httpx.AsyncClient, dim: int) -> None:
+    async def _ensure_collection(self, dim: int) -> None:
         if self._ready:
             return
-        r = await client.get(f"{self.qdrant}/collections/{self.collection}")
+        client = self._client()
+        r = await client.get(
+            f"{self.qdrant}/collections/{self.collection}", timeout=self.store_timeout
+        )
         if r.status_code == 200:
             # Collection existante : si la dimension a changé (modèle d'embedding
             # différent, ex. Qwen3 4096 → nomic 768), on recrée — les anciens vecteurs
@@ -57,30 +73,38 @@ class LongTermMemory:
         await client.put(
             f"{self.qdrant}/collections/{self.collection}",
             json={"vectors": {"size": dim, "distance": "Cosine"}},
+            timeout=self.store_timeout,
         )
         self._ready = True
 
-    async def recall(self, text: str) -> list[str]:
-        """Souvenirs pertinents (texte) ou [] si indisponible."""
+    async def recall(self, text: str) -> tuple[list[str], list[float] | None]:
+        """Souvenirs pertinents + le vecteur de la requête (réutilisable pour `store`).
+
+        Sur le chemin de réponse → timeout serré (`recall_timeout`) : si l'embed ou Qdrant
+        traîne, on dégrade vite en `([], None)` plutôt que de faire patienter le client.
+        Le vecteur renvoyé évite de ré-embedder le même texte au moment du `store`.
+        """
         try:
-            async with httpx.AsyncClient(timeout=20) as client:
-                vec = await self._embed(client, text)
-                r = await client.post(
-                    f"{self.qdrant}/collections/{self.collection}/points/search",
-                    json={"vector": vec, "limit": self.top_k, "with_payload": True},
-                )
-                if r.status_code == 404:  # collection pas encore créée
-                    return []
-                r.raise_for_status()
-                pts = r.json().get("result", [])
-                return [p["payload"]["text"] for p in pts if p.get("payload", {}).get("text")]
+            client = self._client()
+            vec = await self._embed(text, timeout=self.recall_timeout)
+            r = await client.post(
+                f"{self.qdrant}/collections/{self.collection}/points/search",
+                json={"vector": vec, "limit": self.top_k, "with_payload": True},
+                timeout=self.recall_timeout,
+            )
+            if r.status_code == 404:  # collection pas encore créée → rien à rappeler
+                return [], vec
+            r.raise_for_status()
+            pts = r.json().get("result", [])
+            texts = [p["payload"]["text"] for p in pts if p.get("payload", {}).get("text")]
+            return texts, vec
         except Exception:  # noqa: BLE001 — dégrade silencieusement
-            return []
+            return [], None
 
     async def health(self) -> dict:
         """Diagnostic actif : sonde embeddings + Qdrant + collection, sans rien avaler.
 
-        Contrairement à recall/remember (qui dégradent en silence), expose les erreurs
+        Contrairement à recall/store (qui dégradent en silence), expose les erreurs
         pour pouvoir déboguer la mémoire long-terme sans `kubectl exec`.
         """
         out: dict = {
@@ -91,49 +115,60 @@ class LongTermMemory:
             "embed": {"ok": False},
             "qdrant": {"ok": False},
         }
-        async with httpx.AsyncClient(timeout=20) as client:
-            # 1) embeddings
-            try:
-                vec = await self._embed(client, "ping mémoire")
-                out["embed"] = {"ok": True, "dim": len(vec)}
-            except Exception as e:  # noqa: BLE001 — on veut l'erreur
-                out["embed"] = {"ok": False, "error": f"{type(e).__name__}: {e}"}
-            # 2) Qdrant + collection
-            try:
-                r = await client.get(f"{self.qdrant}/collections")
-                r.raise_for_status()
-                names = [c["name"] for c in r.json().get("result", {}).get("collections", [])]
-                exists = self.collection in names
-                qdrant: dict = {"ok": True, "collections": names, "collection_exists": exists}
-                if exists:
-                    cr = await client.get(f"{self.qdrant}/collections/{self.collection}")
-                    if cr.status_code == 200:
-                        qdrant["points_count"] = cr.json().get("result", {}).get("points_count")
-                out["qdrant"] = qdrant
-            except Exception as e:  # noqa: BLE001
-                out["qdrant"] = {"ok": False, "error": f"{type(e).__name__}: {e}"}
+        client = self._client()
+        # 1) embeddings
+        try:
+            vec = await self._embed("ping mémoire", timeout=self.store_timeout)
+            out["embed"] = {"ok": True, "dim": len(vec)}
+        except Exception as e:  # noqa: BLE001 — on veut l'erreur
+            out["embed"] = {"ok": False, "error": f"{type(e).__name__}: {e}"}
+        # 2) Qdrant + collection
+        try:
+            r = await client.get(f"{self.qdrant}/collections", timeout=self.store_timeout)
+            r.raise_for_status()
+            names = [c["name"] for c in r.json().get("result", {}).get("collections", [])]
+            exists = self.collection in names
+            qdrant: dict = {"ok": True, "collections": names, "collection_exists": exists}
+            if exists:
+                cr = await client.get(
+                    f"{self.qdrant}/collections/{self.collection}", timeout=self.store_timeout
+                )
+                if cr.status_code == 200:
+                    qdrant["points_count"] = cr.json().get("result", {}).get("points_count")
+            out["qdrant"] = qdrant
+        except Exception as e:  # noqa: BLE001
+            out["qdrant"] = {"ok": False, "error": f"{type(e).__name__}: {e}"}
         return out
 
-    async def remember(self, session_id: str, text: str) -> None:
+    async def store(
+        self, session_id: str, text: str, vec: list[float] | None = None
+    ) -> None:
+        """Mémorise un tour. Pensé pour tourner **en tâche de fond** (hors chemin de réponse).
+
+        `vec` : si fourni (réutilisé depuis `recall`), on évite un 2ᵉ embed du même texte.
+        Pas de `?wait=true` : on n'attend pas le flush disque de Qdrant — le client a déjà
+        sa réponse, ce point peut être indexé en arrière-plan.
+        """
         try:
-            async with httpx.AsyncClient(timeout=20) as client:
-                vec = await self._embed(client, text)
-                await self._ensure_collection(client, len(vec))
-                await client.put(
-                    f"{self.qdrant}/collections/{self.collection}/points?wait=true",
-                    json={
-                        "points": [
-                            {
-                                "id": str(uuid.uuid4()),
-                                "vector": vec,
-                                "payload": {
-                                    "text": text,
-                                    "session_id": session_id,
-                                    "ts": time.time(),
-                                },
-                            }
-                        ]
-                    },
-                )
+            if vec is None:
+                vec = await self._embed(text, timeout=self.store_timeout)
+            await self._ensure_collection(len(vec))
+            await self._client().put(
+                f"{self.qdrant}/collections/{self.collection}/points",
+                json={
+                    "points": [
+                        {
+                            "id": str(uuid.uuid4()),
+                            "vector": vec,
+                            "payload": {
+                                "text": text,
+                                "session_id": session_id,
+                                "ts": time.time(),
+                            },
+                        }
+                    ]
+                },
+                timeout=self.store_timeout,
+            )
         except Exception:  # noqa: BLE001 — dégrade silencieusement
             pass