diff --git a/.changeset/readbuffer-release-consumed.md b/.changeset/readbuffer-release-consumed.md
new file mode 100644
index 0000000000..667fdee37c
--- /dev/null
+++ b/.changeset/readbuffer-release-consumed.md
@@ -0,0 +1,6 @@
+---
+'@modelcontextprotocol/client': patch
+'@modelcontextprotocol/server': patch
+---
+
+`ReadBuffer` now releases its backing allocation once the last buffered message is consumed, instead of retaining it through an empty `Buffer` view. Long-lived stdio transports no longer pin consumed chunks (pooled or multi-message buffers) until the next append, and the append after a full drain assigns the incoming chunk directly instead of copying through `Buffer.concat`. Fixes #2536
diff --git a/packages/core-internal/src/shared/stdio.ts b/packages/core-internal/src/shared/stdio.ts
index 8bd794b87b..dc0f2da0de 100644
--- a/packages/core-internal/src/shared/stdio.ts
+++ b/packages/core-internal/src/shared/stdio.ts
@@ -31,7 +31,13 @@ export class ReadBuffer {
             }
 
             const line = this._buffer.toString('utf8', 0, index).replace(/\r$/, '');
-            this._buffer = this._buffer.subarray(index + 1);
+            // Drop the buffer entirely once the last message is consumed: an
+            // empty subarray view is truthy and keeps the backing allocation
+            // (which may hold far more than this message, e.g. pooled or
+            // multi-message chunks) alive, and forces the next append through
+            // Buffer.concat.
+            const remainder = this._buffer.subarray(index + 1);
+            this._buffer = remainder.length === 0 ? undefined : remainder;
 
             try {
                 return deserializeMessage(line);
diff --git a/packages/core-internal/test/shared/stdio.test.ts b/packages/core-internal/test/shared/stdio.test.ts
index f8d27a4c1f..d28a89bb74 100644
--- a/packages/core-internal/test/shared/stdio.test.ts
+++ b/packages/core-internal/test/shared/stdio.test.ts
@@ -34,6 +34,37 @@ test('should be reusable after clearing', () => {
     expect(readBuffer.readMessage()).toEqual(testMessage);
 });
 
+test('releases the backing allocation once the last message is consumed', () => {
+    const readBuffer = new ReadBuffer();
+    readBuffer.append(Buffer.from(JSON.stringify(testMessage) + '\n'));
+
+    expect(readBuffer.readMessage()).toEqual(testMessage);
+    // An empty subarray view is truthy and keeps the backing allocation alive
+    // (one chunk can hold several messages, or be a pooled buffer); the
+    // buffer must be dropped entirely instead.
+    expect((readBuffer as unknown as { _buffer?: Buffer })._buffer).toBeUndefined();
+});
+
+test('appends directly after a full drain', () => {
+    const readBuffer = new ReadBuffer();
+    readBuffer.append(Buffer.from(JSON.stringify(testMessage) + '\n'));
+    readBuffer.readMessage();
+
+    readBuffer.append(Buffer.from(JSON.stringify(testMessage) + '\n'));
+    expect(readBuffer.readMessage()).toEqual(testMessage);
+    expect(readBuffer.readMessage()).toBeNull();
+});
+
+test('keeps a partial next message when bytes follow the newline', () => {
+    const readBuffer = new ReadBuffer();
+    readBuffer.append(Buffer.from(JSON.stringify(testMessage) + '\n' + JSON.stringify(testMessage)));
+    expect(readBuffer.readMessage()).toEqual(testMessage);
+
+    readBuffer.append(Buffer.from('\n'));
+    expect(readBuffer.readMessage()).toEqual(testMessage);
+    expect(readBuffer.readMessage()).toBeNull();
+});
+
 describe('non-JSON line filtering', () => {
     test('should skip empty lines', () => {
         const readBuffer = new ReadBuffer();