diff --git a/README.md b/README.md
index 81ff078..9d101bd 100644
--- a/README.md
+++ b/README.md
@@ -78,7 +78,7 @@ FIT also exposes FIT-specific fields not available from TCX/GPX (e.g. `fractiona
 FIT parsing wraps [`fitdecode`](https://github.com/polyvertex/fitdecode), which decodes against Garmin's public FIT SDK profile — field names, types, and units come from that profile, not this package.
 
 - Newer devices sometimes write only `enhanced_altitude`/`enhanced_speed` (and lap `enhanced_avg_speed`/`enhanced_max_speed`); `parse()` coalesces these into `altitude`/`speed` transparently.
-- Messages/fields fitdecode can't resolve against the profile (e.g. proprietary extensions) are dropped.
+- Messages/fields fitdecode can't resolve against the profile (e.g. proprietary extensions) are kept as raw values under `unknown_<n>` names; `parse()`'s canonical columns still exclude them.
 
 ## TCX & GPX files
 
diff --git a/src/activity_parser/parse_fit_file.py b/src/activity_parser/parse_fit_file.py
index 0db8aa8..c7c06bb 100644
--- a/src/activity_parser/parse_fit_file.py
+++ b/src/activity_parser/parse_fit_file.py
@@ -21,13 +21,13 @@ def copy_fit_frames(
     """Yields FIT data frames from a file-like object."""
     processor = fitdecode.StandardUnitsDataProcessor()
     for frame in fitdecode.FitReader(fit_file, processor=processor):
-        if isinstance(frame, fitdecode.FitDataMessage) and frame.mesg_type is not None:
+        if isinstance(frame, fitdecode.FitDataMessage):
             yield frame
 
 
 def frame_to_dict(frame: fitdecode.FitDataMessage) -> dict[str, Any]:
-    """Convert one FIT frame to a dict, dropping unknown fields."""
-    return {field.name: field.value for field in frame.fields if field.field is not None}
+    """Convert one FIT frame to a dict."""
+    return {field.name: field.value for field in frame.fields}
 
 
 def parse_fit_frames(
@@ -81,9 +81,9 @@ def parse_fit(
 ) -> tuple[pd.DataFrame, pd.DataFrame, dict[str, Any]]:
     """Loads a FIT activity into Pandas DataFrames.
 
-    FIT frames and data fields that are marked as 'unknown' by fitdecode are dropped
-    during import. Assumes that the FIT file is all one activity, i.e. chained FIT files
-    will be merged into one set of return values.
+    Message types and fields fitdecode can't resolve against the FIT SDK profile are
+    kept under fitdecode's ``unknown_<num>`` names. Assumes that the FIT file is all one
+    activity, i.e. chained FIT files will be merged into one set of return values.
 
     Args:
         file: File-like or path-like object. A path-like argument ending in ``.gz`` will
diff --git a/tests/test_parse_fit.py b/tests/test_parse_fit.py
index e1bf1d4..d3e5913 100644
--- a/tests/test_parse_fit.py
+++ b/tests/test_parse_fit.py
@@ -28,6 +28,23 @@ def test_parse_fit_records_index(path):
     assert "file_id" in extra
 
 
+def test_parse_fit_keeps_unknown_record_fields():
+    records, _, _ = parse_fit(EDGE_820)
+    assert "unknown_61" in records.columns
+    assert "unknown_66" in records.columns
+
+
+def test_parse_fit_keeps_unknown_messages_in_extra():
+    _, _, extra = parse_fit(EDGE_820)
+    assert "unknown_22" in extra
+
+
+def test_activity_parser_drops_unknown_fields_from_canonical_output():
+    records, laps, _ = ActivityParser().parse(EDGE_820)
+    assert not any(col.startswith("unknown_") for col in records.columns)
+    assert not any(col.startswith("unknown_") for col in laps.columns)
+
+
 @pytest.mark.parametrize("path", [EDGE_820, FENIX_5], ids=lambda p: p.stem)
 def test_parse_fit_numeric_dtypes(path):
     records, _, _ = parse_fit(path)