diff --git a/crates/engine/src/analysis/models.rs b/crates/engine/src/analysis/models.rs
index bad9fd5ff..a584aebbb 100644
--- a/crates/engine/src/analysis/models.rs
+++ b/crates/engine/src/analysis/models.rs
@@ -63,7 +63,7 @@ impl Models {
         self.face_model.lock().ok().and_then(|m| m.clone())
     }
 
-    fn set_active(&self, task: Task, model: Option<Loaded>) {
+    pub(crate) fn set_active(&self, task: Task, model: Option<Loaded>) {
         activate(self.lane(task), &self.face_model, task, model);
     }
 }
@@ -128,6 +128,9 @@ fn kind(task: Task) -> &'static Kind {
     }
 }
 
+/// How long a Roto Brush model may sit unused before its memory is freed.
+const ROTO_MODEL_IDLE: Duration = Duration::from_secs(600);
+
 /// How long a job that needs the Roto Brush model waits for it to load (or download) before it
 /// gives up. A worker that never reports back must not hang the command.
 const ROTO_MODEL_WAIT: Duration = Duration::from_secs(600);
@@ -290,8 +293,35 @@ impl Session {
         Err(format!("{} is still loading: try again when it is ready, or pass wait: true", info.name))
     }
 
+    /// Free a Roto Brush model that has been idle for [`ROTO_MODEL_IDLE`] (and is not under a job). The
+    /// choice stays: it loads again when a job or a stroke needs it.
+    fn unload_idle_roto_model(&mut self) {
+        if self.is_roto_running() || self.models.status(Task::Mask).busy.is_some() {
+            return;
+        }
+        let Some(m) = crate::effects::roto::installed_model() else { return };
+        if m.idle_secs().is_some_and(|s| s >= ROTO_MODEL_IDLE.as_secs_f64()) {
+            let name = m.info().name;
+            drop(m);
+            self.models.set_active(Task::Mask, None);
+            self.events.push(crate::Event::Toast {
+                message: format!("{name} unloaded after {} minutes idle; it loads again when it is used", ROTO_MODEL_IDLE.as_secs() / 60),
+                error: false,
+            });
+        }
+    }
+
+    /// Load the chosen Roto Brush model again if it was unloaded for being idle (before a stroke uses it).
+    pub fn reload_roto_model(&mut self) {
+        let chosen = self.chosen_model(Task::Mask).to_string();
+        if chosen != seg::CLASSICAL && self.models.status(Task::Mask).active.is_none() && self.models.status(Task::Mask).busy.is_none() {
+            self.apply_model(Task::Mask, true);
+        }
+    }
+
     /// Finish what the model workers did: a toast, and loading a model that just arrived.
     pub fn poll_models(&mut self) {
+        self.unload_idle_roto_model();
         for task in [Task::Mask, Task::Face] {
             let done = self.models.lane(task).lock().ok().and_then(|mut s| s.done.take());
             if let Some((message, error)) = done {
diff --git a/crates/engine/src/cmd/roto_cmds.rs b/crates/engine/src/cmd/roto_cmds.rs
index 0931236a2..481a23b5d 100644
--- a/crates/engine/src/cmd/roto_cmds.rs
+++ b/crates/engine/src/cmd/roto_cmds.rs
@@ -196,6 +196,8 @@ fn points_p(p: &Value, cmd: &str) -> Result<Vec<[f64; 2]>> {
 
 fn stroke(s: &mut Session, p: &Value) -> Result<Value> {
     let cmd = "roto.stroke";
+    // A model unloaded for being idle comes back before the stroke is segmented with it.
+    s.reload_roto_model();
     let (cid, lid) = match roto_p(s, p, cmd) {
         Ok((c, l, _)) => (c, l),
         Err(_) => layer_p(s, p, cmd)?,
diff --git a/crates/engine/src/tests/models.rs b/crates/engine/src/tests/models.rs
index 0b714d717..fab025a3a 100644
--- a/crates/engine/src/tests/models.rs
+++ b/crates/engine/src/tests/models.rs
@@ -105,3 +105,40 @@ fn models_are_listed_chosen_and_verified() {
     }
     let _ = std::fs::remove_dir_all(d);
 }
+
+/// A Roto Brush model that has been idle for ten minutes is unloaded (the choice stays), one in use is not,
+/// and a model that does not track use is never unloaded.
+#[test]
+fn an_idle_roto_model_is_unloaded() {
+    use effectcraft_segment::{Loaded, MaskModel, ModelInfo, Prompt};
+    struct Fake(Option<f64>);
+    impl MaskModel for Fake {
+        fn info(&self) -> &'static ModelInfo {
+            &effectcraft_segment::SAM2_TINY
+        }
+        fn segment(&self, _: &[[f32; 3]], _: usize, _: usize, _: &Prompt) -> effectcraft_segment::Result<Vec<f32>> {
+            Err("no weights".into())
+        }
+        fn idle_secs(&self) -> Option<f64> {
+            self.0
+        }
+    }
+    crate::tests_roto::hold_roto_cache_test_lock();
+    let (mut s, d) = session("idle");
+    for (idle, unloaded) in [(Some(5.0), false), (None, false), (Some(601.0), true)] {
+        let m: Arc<dyn MaskModel> = Arc::new(Fake(idle));
+        s.models.set_active(Task::Mask, Some(Loaded::Mask(m)));
+        assert_eq!(s.models.status(Task::Mask).active.as_deref(), Some("sam2.1-hiera-tiny"));
+        s.poll_models();
+        assert_eq!(s.models.status(Task::Mask).active.is_none(), unloaded, "idle {idle:?}");
+        assert_eq!(crate::effects::roto::installed_model().is_none(), unloaded, "the effect's model follows");
+        if unloaded {
+            assert!(s.events.iter().any(|e| matches!(e, crate::Event::Toast { message, .. } if message.contains("unloaded after 10 minutes"))));
+        }
+    }
+    // the choice survives, and with the classic engine chosen reloading does nothing
+    s.reload_roto_model();
+    assert!(s.models.status(Task::Mask).active.is_none());
+    crate::effects::roto::set_model(None);
+    let _ = std::fs::remove_dir_all(d);
+}
diff --git a/crates/sam2/src/mask.rs b/crates/sam2/src/mask.rs
index a0eac9cd6..f917e29ef 100644
--- a/crates/sam2/src/mask.rs
+++ b/crates/sam2/src/mask.rs
@@ -54,11 +54,17 @@ fn fingerprint(rgb: &[[f32; 3]], w: usize, h: usize) -> u64 {
 pub struct Sam2Mask {
     model: Arc<VideoModel>,
     last: Mutex<Option<(u64, Arc<Encoded>)>>,
+    used: Mutex<std::time::Instant>,
 }
 
 impl Sam2Mask {
     pub fn new(model: Arc<VideoModel>) -> Sam2Mask {
-        Sam2Mask { model, last: Mutex::new(None) }
+        Sam2Mask { model, last: Mutex::new(None), used: Mutex::new(std::time::Instant::now()) }
+    }
+
+    /// Note that the model is being used (for [`MaskModel::idle_secs`]).
+    pub(crate) fn touch(&self) {
+        *self.used.lock().unwrap_or_else(PoisonError::into_inner) = std::time::Instant::now();
     }
 
     /// The video model, for trackers built on the same weights.
@@ -103,12 +109,17 @@ impl MaskModel for Sam2Mask {
     }
 
     fn segment(&self, rgb: &[[f32; 3]], w: usize, h: usize, prompt: &effectcraft_segment::Prompt) -> effectcraft_segment::Result<Vec<f32>> {
+        self.touch();
         self.probability(rgb, w, h, prompt).map_err(|e| e.to_string())
     }
 
     fn tracker(&self) -> Option<&dyn effectcraft_segment::ObjectTracker> {
         Some(self)
     }
+
+    fn idle_secs(&self) -> Option<f64> {
+        Some(self.used.lock().unwrap_or_else(PoisonError::into_inner).elapsed().as_secs_f64())
+    }
 }
 
 /// Verify the installed file against the registry and build the model.
diff --git a/crates/sam2/src/track.rs b/crates/sam2/src/track.rs
index 2b06d08cd..2476ddb74 100644
--- a/crates/sam2/src/track.rs
+++ b/crates/sam2/src/track.rs
@@ -151,6 +151,9 @@ impl ObjectTracker for Sam2Mask {
         cancel: &dyn Fn() -> bool,
         out: &mut dyn FnMut(usize, i64, TrackedFrame),
     ) -> effectcraft_segment::Result<()> {
-        self.run_track(req, frames, cancel, out).map_err(|e| e.to_string())
+        self.touch();
+        let r = self.run_track(req, frames, cancel, out).map_err(|e| e.to_string());
+        self.touch();
+        r
     }
 }
diff --git a/crates/segment/src/lib.rs b/crates/segment/src/lib.rs
index 0a9a0494a..bd0b1e672 100644
--- a/crates/segment/src/lib.rs
+++ b/crates/segment/src/lib.rs
@@ -49,6 +49,11 @@ pub trait MaskModel: Send + Sync {
     fn tracker(&self) -> Option<&dyn ObjectTracker> {
         None
     }
+    /// Seconds since the model was last used (`None`: it does not track use, so it is never unloaded
+    /// for being idle). Large models free their memory after a quiet spell and load again on use.
+    fn idle_secs(&self) -> Option<f64> {
+        None
+    }
 }
 
 /// One decoded frame handed to a tracker (straight RGB 0–1, row-major `w×h`).
diff --git a/docs/roto-brush-sam2.md b/docs/roto-brush-sam2.md
index a4e58a36d..e61f49be7 100644
--- a/docs/roto-brush-sam2.md
+++ b/docs/roto-brush-sam2.md
@@ -38,6 +38,13 @@ frames before the corrected frame are not rebuilt, so frames just after it can d
 Tracked frames of a session are in memory; a project reopened in a new process re-tracks the whole span after a stroke
 change (its saved tracking is stale as a whole).
 
+## Memory and limits
+The weights take about 156 MB. A model idle for 10 minutes is unloaded (a toast says so; the choice stays) and loads
+again when a Propagate or a stroke needs it. The tracker holds the memory banks of the objects it follows (about 2 MB per
+stored frame, pruned to the frames later ones can still look at) and refuses to grow past 2 GiB, or past a time limit when
+one is set, with a clear error instead of running out of memory. Two runs of the same prompts give bit-identical masks on
+the CPU.
+
 ## How it works
 `crates/sam2` is a Rust port (on [candle](https://github.com/huggingface/candle), native only) of the Hugging Face
 `transformers` SAM 2 implementation, Apache-2.0 (see its `NOTICE`). Every stage is checked against PyTorch by