diff --git a/crates/polars-python/src/file.rs b/crates/polars-python/src/file.rs
index 741c5c695152..e48111077619 100644
--- a/crates/polars-python/src/file.rs
+++ b/crates/polars-python/src/file.rs
@@ -322,14 +322,20 @@ pub fn get_python_scan_source_input(
     write: bool,
 ) -> PyResult<PythonScanSourceInput> {
     Python::with_gil(|py| {
-        let py_f_0 = py_f;
-        let py_f = py_f_0.clone_ref(py).into_bound(py);
+        let py_f = py_f.into_bound(py);
+
+        // CPython has some internal tricks that means much of the time
+        // BytesIO.getvalue() involves no memory copying, unlike
+        // BytesIO.read(). So we want to handle BytesIO specially in order
+        // to save memory.
+        let py_f = read_if_bytesio(py_f);
 
         // If the pyobject is a `bytes` class
         if let Ok(b) = py_f.downcast::<PyBytes>() {
             return Ok(PythonScanSourceInput::Buffer(MemSlice::from_arc(
                 b.as_bytes(),
-                Arc::new(py_f_0),
+                // We want to specifically keep alive the PyBytes object.
+                Arc::new(b.clone().unbind()),
             )));
         }
 
@@ -373,15 +379,17 @@ pub fn get_file_like(f: PyObject, truncate: bool) -> PyResult<Box<dyn FileLike>>
     Ok(get_either_file(f, truncate)?.into_dyn())
 }
 
-/// If the give file-like is a BytesIO, read its contents.
+/// If the give file-like is a BytesIO, read its contents in a memory-efficient
+/// way.
 fn read_if_bytesio(py_f: Bound<PyAny>) -> Bound<PyAny> {
-    if py_f.getattr("read").is_ok() {
+    let bytes_io = py_f.py().import("io").unwrap().getattr("BytesIO").unwrap();
+    if py_f.is_instance(&bytes_io).unwrap() {
+        // Note that BytesIO has some memory optimizations ensuring that much of
+        // the time getvalue() doesn't need to copy the underlying data:
         let Ok(bytes) = py_f.call_method0("getvalue") else {
             return py_f;
         };
-        if bytes.downcast::<PyBytes>().is_ok() || bytes.downcast::<PyString>().is_ok() {
-            return bytes.clone();
-        }
+        return bytes;
     }
     py_f
 }
diff --git a/py-polars/src/memory.rs b/py-polars/src/memory.rs
index e7abf2b7d51c..70b0ceb8ac8c 100644
--- a/py-polars/src/memory.rs
+++ b/py-polars/src/memory.rs
@@ -69,6 +69,7 @@ unsafe impl<A: GlobalAlloc> GlobalAlloc for TracemallocAllocator<A> {
     }
 
     unsafe fn realloc(&self, ptr: *mut u8, layout: std::alloc::Layout, new_size: usize) -> *mut u8 {
+        PyTraceMalloc_Untrack(TRACEMALLOC_DOMAIN, ptr as uintptr_t);
         let result = self.wrapped_alloc.realloc(ptr, layout, new_size);
         PyTraceMalloc_Track(TRACEMALLOC_DOMAIN, result as uintptr_t, new_size);
         result
diff --git a/py-polars/tests/unit/conftest.py b/py-polars/tests/unit/conftest.py
index 8197872831b5..11af90824728 100644
--- a/py-polars/tests/unit/conftest.py
+++ b/py-polars/tests/unit/conftest.py
@@ -5,6 +5,7 @@
 import random
 import string
 import sys
+import time
 import tracemalloc
 from typing import TYPE_CHECKING, Any, cast
 
@@ -205,7 +206,11 @@ def get_peak(self) -> int:
         return tracemalloc.get_traced_memory()[1]
 
 
-@pytest.fixture
+# The bizarre syntax is from
+# https://github.com/pytest-dev/pytest/issues/1368#issuecomment-2344450259 - we
+# need to mark any test using this fixture as slow because we have a sleep
+# added to work around a CPython bug, see the end of the function.
+@pytest.fixture(params=[pytest.param(0, marks=pytest.mark.slow)])
 def memory_usage_without_pyarrow() -> Generator[MemoryUsage, Any, Any]:
     """
     Provide an API for measuring peak memory usage.
@@ -231,6 +236,10 @@ def memory_usage_without_pyarrow() -> Generator[MemoryUsage, Any, Any]:
     try:
         yield MemoryUsage()
     finally:
+        # Workaround for https://github.com/python/cpython/issues/128679
+        time.sleep(1)
+        gc.collect()
+
         tracemalloc.stop()
 
 
diff --git a/py-polars/tests/unit/io/test_scan.py b/py-polars/tests/unit/io/test_scan.py
index 569c27260513..4f19698488f2 100644
--- a/py-polars/tests/unit/io/test_scan.py
+++ b/py-polars/tests/unit/io/test_scan.py
@@ -16,6 +16,7 @@
 
 if TYPE_CHECKING:
     from polars._typing import SchemaDict
+    from tests.unit.conftest import MemoryUsage
 
 
 @dataclass
@@ -929,3 +930,30 @@ def test_predicate_stats_eval_nested_binary() -> None:
         ),
         pl.DataFrame({"x": [2]}),
     )
+
+
+@pytest.mark.slow
+@pytest.mark.parametrize("streaming", [True, False])
+def test_scan_csv_bytesio_memory_usage(
+    streaming: bool,
+    memory_usage_without_pyarrow: MemoryUsage,
+) -> None:
+    memory_usage = memory_usage_without_pyarrow
+
+    # Create CSV that is ~70-85 MB in size:
+    f = io.BytesIO()
+    df = pl.DataFrame({"mydata": pl.int_range(0, 10_000_000, eager=True)})
+    df.write_csv(f)
+    assert 70_000_000 < f.tell() < 85_000_000
+    f.seek(0, 0)
+
+    # A lazy scan shouldn't make a full copy of the data:
+    starting_memory = memory_usage.get_current()
+    assert (
+        pl.scan_csv(f)
+        .filter(pl.col("mydata") == 9_999_999)
+        .collect(new_streaming=streaming)  # type: ignore[call-overload]
+        .item()
+        == 9_999_999
+    )
+    assert memory_usage.get_peak() - starting_memory < 10_000_000