diff --git a/benchmarks/graphiti/src/memory.py b/benchmarks/graphiti/src/memory.py
index 3d16eb0..29ee1dd 100644
--- a/benchmarks/graphiti/src/memory.py
+++ b/benchmarks/graphiti/src/memory.py
@@ -32,6 +32,9 @@
 with the same models. Override with ``--param model=...`` /
 ``--param embedding_model=...``; an explicit ``--param model=none`` (or
 ``embedding_model=none``) falls back to graphiti-core's own default.
+The LLM client runs with a raised completion ceiling (``LLM_MAX_TOKENS``)
+so gpt-5-mini's reasoning bursts cannot truncate structured output
+mid-JSON, and teardown closes every async client before its event loop.
 """
 
 import asyncio