diff --git a/frontend/src/components/cortex/cortex-app.tsx b/frontend/src/components/cortex/cortex-app.tsx
index 2712b8c..e2925ea 100644
--- a/frontend/src/components/cortex/cortex-app.tsx
+++ b/frontend/src/components/cortex/cortex-app.tsx
@@ -150,6 +150,7 @@ import {
 import { Onboarding } from "./onboarding";
 import { CaptureModal } from "./capture";
 import { Markdown } from "./markdown";
+import { StreamingAnswer } from "./streaming-answer";
 import { MediaBlock } from "./media-block";
 import { SourceChips, type SourceItem } from "./sources";
 
@@ -3885,7 +3886,11 @@ export function CortexApp({
                             <MediaBlock media={m.media} />
                           ) : (
                             <div className="atext">
-                              {m.streaming ? m.a : <Markdown text={m.a} />}
+                              {m.streaming ? (
+                                <StreamingAnswer />
+                              ) : (
+                                <Markdown text={m.a} />
+                              )}
                             </div>
                           )}
                           {!m.streaming && !m.media && (
diff --git a/frontend/src/components/cortex/streaming-answer.tsx b/frontend/src/components/cortex/streaming-answer.tsx
new file mode 100644
index 0000000..8ed8693
--- /dev/null
+++ b/frontend/src/components/cortex/streaming-answer.tsx
@@ -0,0 +1,12 @@
+"use client";
+
+import { useStreamStore } from "@/lib/cortex/stream-store";
+
+// Renders the in-flight answer text while it streams. It subscribes ONLY to the
+// stream store, so each typewriter tick re-renders this node alone rather than the
+// whole app. Once streaming finishes the chat message holds the final text and is
+// rendered as Markdown instead.
+export function StreamingAnswer() {
+  const text = useStreamStore((s) => s.text);
+  return <>{text}</>;
+}
diff --git a/frontend/src/lib/cortex/store.ts b/frontend/src/lib/cortex/store.ts
index 7649575..4be252b 100644
--- a/frontend/src/lib/cortex/store.ts
+++ b/frontend/src/lib/cortex/store.ts
@@ -1,5 +1,6 @@
 "use client";
 import { create } from "zustand";
+import { useStreamStore } from "./stream-store";
 import {
   type Memory,
   type CortexEvent,
@@ -999,16 +1000,17 @@ export const useCortex = create<State>((set, get) => ({
       activeId: get().activeId,
     });
 
-    // Reveal the answer progressively, but cap the work: each store write
-    // re-renders the whole app tree (the top-level view subscribes to the store),
-    // so a per-word/30ms tick meant ~33 full re-renders per second. Reveal several
-    // words per tick and bound the run to STREAM_MAX_TICKS updates regardless of
-    // length, keeping the typewriter feel at a fraction of the render cost.
+    // Reveal the answer progressively. Each per-tick partial goes into the dedicated
+    // stream store (see stream-store.ts), NOT the main store, so only the small
+    // <StreamingAnswer /> node re-renders per tick instead of the whole app tree
+    // (the top-level view subscribes to the entire main store). The final text is
+    // committed back to the chat ONCE, which is the only main-store write of the run.
     const STREAM_MAX_TICKS = 24;
     const STREAM_INTERVAL_MS = 55;
     const stream = (text: string) => {
       const words = text.split(" ");
       const perTick = Math.max(1, Math.ceil(words.length / STREAM_MAX_TICKS));
+      useStreamStore.getState().setText("");
       let i = 0;
       const tick = setInterval(() => {
         if (i >= words.length) {
@@ -1022,6 +1024,7 @@ export const useCortex = create<State>((set, get) => ({
             }
             return { chat };
           });
+          useStreamStore.getState().setText("");
           persist({
             memories: get().memories,
             events: get().events,
@@ -1033,13 +1036,7 @@ export const useCortex = create<State>((set, get) => ({
           return;
         }
         i = Math.min(words.length, i + perTick);
-        const partial = words.slice(0, i).join(" ");
-        set((s) => {
-          const chat = [...s.chat];
-          const last = chat[chat.length - 1];
-          if (last) last.a = partial;
-          return { chat };
-        });
+        useStreamStore.getState().setText(words.slice(0, i).join(" "));
       }, STREAM_INTERVAL_MS);
     };
 
diff --git a/frontend/src/lib/cortex/stream-store.ts b/frontend/src/lib/cortex/stream-store.ts
new file mode 100644
index 0000000..2a8d75d
--- /dev/null
+++ b/frontend/src/lib/cortex/stream-store.ts
@@ -0,0 +1,17 @@
+import { create } from "zustand";
+
+// The answer typewriter reveals text in ~24 ticks. Writing each tick into the main
+// store re-rendered the whole app tree (the top-level view subscribes to the entire
+// store), which is the biggest source of jank during a reply. Park the in-flight
+// partial text here instead: only the small <StreamingAnswer /> reads it, so a tick
+// re-renders that node alone. The final text is committed back to the chat once,
+// streaming is done.
+interface StreamState {
+  text: string;
+  setText: (text: string) => void;
+}
+
+export const useStreamStore = create<StreamState>((set) => ({
+  text: "",
+  setText: (text) => set({ text }),
+}));