diff --git a/litellm/responses/compaction.py b/litellm/responses/compaction.py
index 597ec8cf327..75b9b3bcd19 100644
--- a/litellm/responses/compaction.py
+++ b/litellm/responses/compaction.py
@@ -5,6 +5,7 @@ Summarizes conversation history when input tokens exceed a configured threshold,
 using the same model the caller is already using. Works across all providers.
 """
 
+import copy
 import re
 from typing import Any, Dict, List, Optional, Tuple
 
@@ -127,7 +128,7 @@ async def maybe_compact_context(
     if custom_llm_provider is not None:
         acompletion_kwargs["custom_llm_provider"] = custom_llm_provider
     if litellm_metadata is not None:
-        acompletion_kwargs["metadata"] = {**litellm_metadata}
+        acompletion_kwargs["metadata"] = copy.deepcopy(litellm_metadata)
 
     summary_response = await litellm.acompletion(**acompletion_kwargs)