diff --git a/Cargo.toml b/Cargo.toml
index b38c520..9578ed3 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -14,8 +14,13 @@ categories = ["algorithms", "mathematics", "science"]
 num-traits = "0.2.18"
 kdtree = "0.7.0"
 rayon = { version = "1.10.0", optional = true }
+serde = { version = "1.0", features = ["derive"], optional = true }
+
+[dev-dependencies]
+bincode = "1.3"
 
 [features]
 default = ["serial"]
 serial = []
 parallel = ["dep:rayon"]
+serde = ["dep:serde"]
diff --git a/src/data_wrappers.rs b/src/data_wrappers.rs
index a7f529d..321f49e 100644
--- a/src/data_wrappers.rs
+++ b/src/data_wrappers.rs
@@ -12,9 +12,25 @@ pub(crate) struct SLTNode<T> {
     pub(crate) size: usize,
 }
 
-pub(crate) struct CondensedNode<T> {
-    pub(crate) node_id: usize,
-    pub(crate) parent_node_id: usize,
-    pub(crate) lambda_birth: T,
-    pub(crate) size: usize,
+/// A node in the condensed cluster tree produced by HDBSCAN. Exposed
+/// publicly to enable external `approximate_predict`-style inference on
+/// new points (assigning previously-unseen samples to existing clusters
+/// without re-running the full algorithm).
+///
+/// Marked `#[non_exhaustive]` so additional fields can be added in
+/// future revisions without breaking external consumers.
+#[non_exhaustive]
+#[cfg_attr(feature = "serde", derive(serde::Serialize, serde::Deserialize))]
+#[cfg_attr(
+    feature = "serde",
+    serde(bound(
+        serialize = "T: serde::Serialize",
+        deserialize = "T: serde::Deserialize<'de>"
+    ))
+)]
+pub struct CondensedNode<T> {
+    pub node_id: usize,
+    pub parent_node_id: usize,
+    pub lambda_birth: T,
+    pub size: usize,
 }
diff --git a/src/hdbscan.rs b/src/hdbscan.rs
index 0155c0f..695fc60 100644
--- a/src/hdbscan.rs
+++ b/src/hdbscan.rs
@@ -15,7 +15,11 @@ use num_traits::Float;
 use std::collections::{HashMap, VecDeque};
 use std::ops::Range;
 
-type CondensedTree<T> = Vec<CondensedNode<T>>;
+/// The condensed cluster tree produced internally by HDBSCAN, used to
+/// derive cluster labels. Exposed publicly to enable external
+/// `approximate_predict`-style inference; obtain an instance via
+/// [`Hdbscan::cluster_with_tree`].
+pub type CondensedTree<T> = Vec<CondensedNode<T>>;
 
 /// The HDBSCAN clustering algorithm in Rust. Generic over floating point numeric types.
 #[derive(Debug, Clone, PartialEq)]
@@ -65,6 +69,29 @@ impl<T: Float> Hdbscan<'_, T> {
     ///assert_eq!(-1, labels[10]);
     /// ```
     pub fn cluster(&self) -> Result<Vec<i32>, HdbscanError> {
+        self.cluster_internal().map(|(labels, _tree)| labels)
+    }
+
+    /// Same as [`cluster`](Hdbscan::cluster), but additionally returns
+    /// the condensed cluster tree used internally to derive the cluster
+    /// labels. The condensed tree is required for
+    /// `approximate_predict`-style inference: assigning previously-unseen
+    /// points to existing clusters without re-running the full algorithm.
+    ///
+    /// # Returns
+    /// * A `Result` whose `Ok` contains `(labels, condensed_tree)`.
+    ///   Labels follow the same semantics as [`cluster`](Hdbscan::cluster);
+    ///   the condensed tree is a `Vec<CondensedNode<T>>` describing the
+    ///   cluster hierarchy.
+    pub fn cluster_with_tree(
+        &self,
+    ) -> Result<(Vec<i32>, CondensedTree<T>), HdbscanError> {
+        self.cluster_internal()
+    }
+
+    fn cluster_internal(
+        &self,
+    ) -> Result<(Vec<i32>, CondensedTree<T>), HdbscanError> {
         DataValidator::new(self.data, &self.hp).validate_input_data()?;
 
         let core_dist_calculator = CoreDistanceCalculator::new(self.data, &self.hp);
@@ -79,7 +106,7 @@ impl<T: Float> Hdbscan<'_, T> {
         let winning_clusters = self.extract_winning_clusters(&condensed_tree);
         let labelled_data = self.label_data(&winning_clusters, &condensed_tree);
 
-        Ok(labelled_data)
+        Ok((labelled_data, condensed_tree))
     }
 }
 
@@ -806,3 +833,36 @@ impl<'a, T: Float> Hdbscan<'a, T> {
         }
     }
 }
+
+#[cfg(all(test, feature = "serde"))]
+mod serde_tests {
+    use super::CondensedTree;
+    use crate::data_wrappers::CondensedNode;
+
+    #[test]
+    fn condensed_tree_bincode_roundtrip() {
+        let tree: CondensedTree<f64> = vec![
+            CondensedNode {
+                node_id: 5,
+                parent_node_id: 11,
+                lambda_birth: 0.42,
+                size: 3,
+            },
+            CondensedNode {
+                node_id: 6,
+                parent_node_id: 11,
+                lambda_birth: 0.42,
+                size: 4,
+            },
+        ];
+        let bytes = bincode::serialize(&tree).expect("serialize tree");
+        let decoded: CondensedTree<f64> =
+            bincode::deserialize(&bytes).expect("deserialize tree");
+        assert_eq!(decoded.len(), 2);
+        assert_eq!(decoded[0].node_id, 5);
+        assert_eq!(decoded[0].parent_node_id, 11);
+        assert_eq!(decoded[0].lambda_birth, 0.42);
+        assert_eq!(decoded[0].size, 3);
+        assert_eq!(decoded[1].node_id, 6);
+    }
+}
diff --git a/src/lib.rs b/src/lib.rs
index 151f4f1..6391985 100644
--- a/src/lib.rs
+++ b/src/lib.rs
@@ -51,9 +51,10 @@
 
 pub use crate::centers::Center;
 pub use crate::core_distances::NnAlgorithm;
+pub use crate::data_wrappers::CondensedNode;
 pub use crate::distance::DistanceMetric;
 pub use crate::error::HdbscanError;
-pub use crate::hdbscan::Hdbscan;
+pub use crate::hdbscan::{CondensedTree, Hdbscan};
 pub use crate::hyper_parameters::{HdbscanHyperParams, HyperParamBuilder};
 
 mod centers;