diff --git a/tests/features/test_model_lifecycle.py b/tests/features/test_model_lifecycle.py
new file mode 100644
index 000000000..f2c99866a
--- /dev/null
+++ b/tests/features/test_model_lifecycle.py
@@ -0,0 +1,149 @@
+"""Test model lifecycle (load/unload/reload)
+
+Tests for Issue: Verify that models can be cleanly unloaded and reloaded,
+with proper cleanup of variable references to the shared library.
+"""
+import pytest
+import numpy as np
+from pygenn import (
+    init_weight_update,
+    init_postsynaptic,
+    init_sparse_connectivity,
+    create_custom_update_model,
+    create_var_ref,
+    types,
+)
+
+
+def _check_vars_cleared(vars_dict, component_name, expected_none=True):
+    """Helper to verify all variables in a dict are (or aren't) cleared.
+
+    Args:
+        vars_dict: Dictionary of variables to check
+        component_name: Name of component for error messages
+        expected_none: If True, assert vars are None; if False, assert not None
+    """
+    for var_name, var in vars_dict.items():
+        if expected_none:
+            assert var._view is None, f"{component_name}.vars[{var_name}]._view not cleared"
+            assert var._array is None, f"{component_name}.vars[{var_name}]._array not cleared"
+        else:
+            assert var._view is not None, f"{component_name}.vars[{var_name}]._view is None"
+            assert var._array is not None, f"{component_name}.vars[{var_name}]._array is None"
+
+
+@pytest.mark.parametrize("precision", [types.Double, types.Float])
+def test_model_load_unload_cycle(make_model, backend, precision):
+    """Test that models can be unloaded and reloaded correctly.
+
+    This test creates a comprehensive model with:
+    - Neuron populations (pre and post)
+    - Synapse populations with sparse connectivity (STDP)
+    - Current sources
+    - Custom updates
+
+    Then verifies:
+    1. Model loads and runs correctly
+    2. After unload, all variable views/arrays are None
+    3. Model can be reloaded
+    4. Reloaded model runs correctly
+    """
+    # Create comprehensive model
+    model = make_model(precision, "test_lifecycle", backend)
+
+    # LIF neuron parameters
+    lif_params = {
+        "C": 1.0,
+        "TauM": 20.0,
+        "Vrest": -70.0,
+        "Vreset": -70.0,
+        "Vthresh": -50.0,
+        "Ioffset": 0.0,
+        "TauRefrac": 5.0,
+    }
+    lif_init = {"V": -70.0, "RefracTime": 0.0}
+
+    # Add neuron populations
+    pre = model.add_neuron_population("pre", 10, "LIF", lif_params, lif_init)
+    post = model.add_neuron_population("post", 10, "LIF", lif_params, lif_init)
+
+    # Add synapses with STDP
+    stdp_params = {
+        "tauPlus": 20.0,
+        "tauMinus": 20.0,
+        "Aplus": 0.01,
+        "Aminus": 0.01,
+        "Wmin": 0.0,
+        "Wmax": 1.0,
+    }
+    stdp_init = {"g": 0.5}
+
+    syn = model.add_synapse_population(
+        "synapses",
+        "SPARSE",
+        pre,
+        post,
+        init_weight_update("STDP", stdp_params, stdp_init),
+        init_postsynaptic("DeltaCurr", {}),
+        init_sparse_connectivity("FixedProbability", {"prob": 0.5}),
+    )
+
+    # Add current source
+    cs = model.add_current_source("current", "DC", pre, {"amp": 5.0}, {})
+
+    # Add custom update
+    cu_model = create_custom_update_model(
+        "test_update", update_code="V *= 0.99;", var_refs=[("V", "scalar")]
+    )
+    model.add_custom_update(
+        "TestUpdate",
+        "CustomUpdate",
+        cu_model,
+        {},
+        {},
+        var_refs={"V": create_var_ref(pre, "V")},
+    )
+
+    # Build model
+    model.build()
+
+    # ========== FIRST LOAD ==========
+    model.load()
+    assert model._loaded == True
+
+    # Verify variables are loaded (have views/arrays)
+    assert pre.vars["V"]._view is not None
+    assert pre.vars["V"]._array is not None
+
+    # Run simulation
+    for _ in range(10):
+        model.step_time()
+    model.custom_update("CustomUpdate")
+
+    # ========== UNLOAD ==========
+    model.unload()
+
+    # Verify model is unloaded
+    assert model._loaded == False
+    assert model._runtime is None
+
+    # CRITICAL: Verify all variable references are cleared
+    # (This is what prevents the shared library from unloading properly)
+    _check_vars_cleared(pre.vars, "pre")
+    _check_vars_cleared(post.vars, "post")
+    _check_vars_cleared(syn.vars, "synapses")
+    _check_vars_cleared(cs.vars, "current_source")
+    _check_vars_cleared(model.custom_updates["TestUpdate"].vars, "custom_update")
+
+    # ========== RELOAD ==========
+    model.load()
+    assert model._loaded == True
+
+    # Verify variables are reloaded (have views/arrays again)
+    _check_vars_cleared(pre.vars, "pre", expected_none=False)
+    _check_vars_cleared(post.vars, "post", expected_none=False)
+
+    # Run simulation again to verify reloaded model works
+    for _ in range(10):
+        model.step_time()
+    model.custom_update("CustomUpdate")