diff --git a/Cargo.lock b/Cargo.lock
index fdaa928aed..4117b73f8d 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -155,6 +155,7 @@ dependencies = [
  "agentos-native-sidecar-core",
  "agentos-runtime",
  "agentos-sidecar-protocol",
+ "agentos-v8-runtime",
  "agentos-vfs",
  "agentos-vfs-core",
  "agentos-vm-config",
diff --git a/crates/native-sidecar/Cargo.toml b/crates/native-sidecar/Cargo.toml
index 120a51a87c..4fca8a80ad 100644
--- a/crates/native-sidecar/Cargo.toml
+++ b/crates/native-sidecar/Cargo.toml
@@ -21,6 +21,7 @@ agentos-sidecar-protocol = { workspace = true }
 agentos-native-sidecar-core = { workspace = true }
 agentos-runtime = { workspace = true }
 agentos-execution = { workspace = true }
+agentos-v8-runtime = { workspace = true }
 agentos-vm-config = { workspace = true }
 agentos-vfs = { workspace = true }
 async-trait = "0.1"
diff --git a/crates/native-sidecar/src/lib.rs b/crates/native-sidecar/src/lib.rs
index fa357a142d..d604c6fb43 100644
--- a/crates/native-sidecar/src/lib.rs
+++ b/crates/native-sidecar/src/lib.rs
@@ -16,6 +16,7 @@ pub(crate) mod json_rpc;
 pub(crate) mod language_execution;
 pub mod limits;
 pub(crate) mod metadata;
+pub(crate) mod native_heap;
 pub(crate) mod ownership_coordinator;
 pub mod package_projection;
 pub(crate) mod plugins;
diff --git a/crates/native-sidecar/src/native_heap.rs b/crates/native-sidecar/src/native_heap.rs
new file mode 100644
index 0000000000..548160dae3
--- /dev/null
+++ b/crates/native-sidecar/src/native_heap.rs
@@ -0,0 +1,51 @@
+//! Return native heap memory to the OS after VM disposal.
+//!
+//! V8 isolates allocate through glibc malloc, which keeps freed arena pages
+//! mapped after an isolate is destroyed. A disposed VM therefore leaves a high
+//! RSS watermark unless the sidecar asks glibc to release free pages.
+
+use std::sync::atomic::{AtomicBool, Ordering};
+use std::time::Duration;
+
+use agentos_runtime::{BlockingJobError, RuntimeContext, TaskClass};
+
+/// Coalescing window. Every disposal inside the window shares one trim, so VM
+/// churn causes at most one trim per window across the whole process.
+const TRIM_COALESCE_WINDOW: Duration = Duration::from_secs(1);
+
+static TRIM_SCHEDULED: AtomicBool = AtomicBool::new(false);
+
+/// Schedule one coalesced native heap trim on the shared blocking executor.
+///
+/// A no-op unless the sidecar runs on Linux with glibc.
+pub(crate) fn schedule_trim_after_vm_disposal(runtime: &RuntimeContext) {
+    if !cfg!(all(target_os = "linux", target_env = "gnu")) {
+        return;
+    }
+    if TRIM_SCHEDULED.swap(true, Ordering::AcqRel) {
+        return;
+    }
+    let blocking = runtime.blocking().clone();
+    let spawned = runtime.spawn(TaskClass::Timer, async move {
+        tokio::time::sleep(TRIM_COALESCE_WINDOW).await;
+        // Clear before the trim so a disposal that finishes during the trim
+        // schedules a later one instead of being absorbed by this one.
+        TRIM_SCHEDULED.store(false, Ordering::Release);
+        match blocking
+            .run(0, agentos_v8_runtime::native_heap::release_free_native_heap)
+            .await
+        {
+            Ok(_) => {}
+            Err(BlockingJobError::ShuttingDown) => {
+                tracing::debug!("skipped native heap trim: sidecar runtime is shutting down");
+            }
+            Err(error) => {
+                tracing::warn!(%error, "ERR_AGENTOS_NATIVE_HEAP_TRIM: native heap trim failed");
+            }
+        }
+    });
+    if let Err(error) = spawned {
+        TRIM_SCHEDULED.store(false, Ordering::Release);
+        tracing::warn!(%error, "ERR_AGENTOS_NATIVE_HEAP_TRIM: failed to schedule native heap trim");
+    }
+}
diff --git a/crates/native-sidecar/src/service.rs b/crates/native-sidecar/src/service.rs
index 84bcf3327f..517d154b79 100644
--- a/crates/native-sidecar/src/service.rs
+++ b/crates/native-sidecar/src/service.rs
@@ -1654,7 +1654,8 @@ where
     /// reset) erroring out with `?` can never strand these maps for the rest of
     /// the process lifetime (H1). This also reclaims extension output buffers,
     /// which was previously removed only on a successful handoff and leaked on VM
-    /// or session disposal (M6).
+    /// or session disposal (M6). Finally, it schedules a coalesced native heap
+    /// trim so the disposed VM's freed V8 memory returns to the OS.
     pub(crate) fn reclaim_vm_tracking(&mut self, session_id: &str, vm_id: &str) {
         self.prune_extension_vm_resource(vm_id);
         self.extension_process_output_buffers
@@ -1662,6 +1663,9 @@ where
         if let Some(session) = self.sessions.get_mut(session_id) {
             session.vm_ids.remove(vm_id);
         }
+        if let Some(runtime_context) = self.runtime_context.as_ref() {
+            crate::native_heap::schedule_trim_after_vm_disposal(runtime_context);
+        }
     }
 
     /// Probe buffered extension process output without retaining the sidecar
diff --git a/crates/native-sidecar/tests/service.rs b/crates/native-sidecar/tests/service.rs
index 2f6a045c62..743afc0eca 100644
--- a/crates/native-sidecar/tests/service.rs
+++ b/crates/native-sidecar/tests/service.rs
@@ -30,6 +30,8 @@ mod limits;
 #[allow(dead_code)]
 #[path = "../src/metadata/mod.rs"]
 mod metadata;
+#[path = "../src/native_heap.rs"]
+mod native_heap;
 #[allow(dead_code)]
 #[path = "../src/package_projection.rs"]
 mod package_projection;
diff --git a/crates/v8-runtime/src/lib.rs b/crates/v8-runtime/src/lib.rs
index f788f0dfd1..2ecf34dcee 100644
--- a/crates/v8-runtime/src/lib.rs
+++ b/crates/v8-runtime/src/lib.rs
@@ -5,6 +5,7 @@ pub mod host_call;
 pub mod ipc;
 pub mod ipc_binary;
 pub mod isolate;
+pub mod native_heap;
 pub mod runtime_protocol;
 pub mod session;
 pub mod snapshot;
diff --git a/crates/v8-runtime/src/native_heap.rs b/crates/v8-runtime/src/native_heap.rs
new file mode 100644
index 0000000000..d68032885d
--- /dev/null
+++ b/crates/v8-runtime/src/native_heap.rs
@@ -0,0 +1,18 @@
+//! Native heap maintenance for memory freed by V8 and other C/C++ code.
+
+/// Return free glibc heap pages to the OS.
+///
+/// V8 isolates and other native code allocate through glibc malloc, which keeps
+/// freed arena memory mapped after an isolate is destroyed. Returns whether
+/// any memory was released. On other platforms and C libraries this is a no-op.
+#[cfg(all(target_os = "linux", target_env = "gnu"))]
+pub fn release_free_native_heap() -> bool {
+    // SAFETY: `malloc_trim` has no preconditions; it takes the allocator's own
+    // arena locks and only releases pages that are already free.
+    unsafe { libc::malloc_trim(0) != 0 }
+}
+
+#[cfg(not(all(target_os = "linux", target_env = "gnu")))]
+pub fn release_free_native_heap() -> bool {
+    false
+}
diff --git a/docs/content/docs/resource-limits.mdx b/docs/content/docs/resource-limits.mdx
index b7eb6b4a7a..c087b15d58 100644
--- a/docs/content/docs/resource-limits.mdx
+++ b/docs/content/docs/resource-limits.mdx
@@ -71,6 +71,18 @@ Library VMs share a sidecar process by default. Limits that apply to the whole
 process, such as `runtime.executor.maxActiveVms`, are set when creating a
 dedicated sidecar with `AgentOs.createSidecar()`, as in the example above.
 
+### Sidecar memory
+
+On Linux with glibc, the sidecar returns memory freed by a disposed VM's V8
+isolates to the OS shortly after disposal.
+
+The sidecar also keeps a pool of up to four pre-started V8 isolates so that
+JavaScript and Python executions start faster. The pool stays resident while
+the sidecar runs, at about 20 MiB per isolate. To cap it, set
+`AGENTOS_V8_WARM_ISOLATES` to a value from `0` to `4` in the environment that
+starts the sidecar. `0` disables the pool and trades memory for slower
+execution starts.
+
 ## Embedded API
 
 Limits apply to VMs created with `AgentOs.create()` and sidecars created with