diff --git a/cli.py b/cli.py
index 86875bcb60cd5..c43ec89407c1c 100644
--- a/cli.py
+++ b/cli.py
@@ -9835,6 +9835,8 @@ def chat(self, message, images: list = None) -> Optional[str]:
             request_overrides=turn_route.get("request_overrides"),
         ):
             return None
+        if getattr(self, "_suppress_single_query_post_turn_prefetch", False):
+            setattr(self.agent, "_suppress_post_turn_prefetch", True)
         
         # Route image attachments based on the active model's vision capability.
         # "native" → pass pixels as OpenAI-style content parts (adapters
@@ -13566,6 +13568,7 @@ def _signal_handler_q(signum, frame):
             sys.exit(1)
         try:
             query, single_query_images = _collect_query_images(query, image)
+            cli._suppress_single_query_post_turn_prefetch = True
             # Kanban workers spawn with ``hermes chat -q "work kanban task <id>"``;
             # the actual task description lives in the task body. Mirror the
             # gateway/CLI behaviour for inbound images by scanning the body for
@@ -13673,6 +13676,11 @@ def _signal_handler_q(signum, frame):
                     ):
                         cli.agent.quiet_mode = True
                         cli.agent.suppress_status_output = True
+                        # Non-interactive single-query runs exit immediately after
+                        # this turn, so there is no next turn to warm.  Suppress
+                        # post-turn memory prefetch to avoid spawning Honcho
+                        # background threads that can race interpreter shutdown.
+                        cli.agent._suppress_post_turn_prefetch = True
                         # Suppress streaming display callbacks so stdout stays
                         # machine-readable (no styled "Hermes" box, no tool-gen
                         # status lines).  The response is printed once below.
diff --git a/hermes_cli/cli_agent_setup_mixin.py b/hermes_cli/cli_agent_setup_mixin.py
index 1041e8fd0b5e8..ad82bbde0b57c 100644
--- a/hermes_cli/cli_agent_setup_mixin.py
+++ b/hermes_cli/cli_agent_setup_mixin.py
@@ -391,9 +391,19 @@ def _init_agent(self, *, model_override: str = None, runtime_override: dict = No
                 notice_callback=self._on_notice,
                 notice_clear_callback=self._on_notice_clear,
             )
-            # Store reference for atexit memory provider shutdown
-            global _active_agent_ref
-            _active_agent_ref = self.agent
+            # Store reference for atexit memory provider shutdown.
+            # This mixin lives in ``hermes_cli.cli_agent_setup_mixin``, but
+            # cleanup reads ``_active_agent_ref`` from the module that owns the
+            # concrete HermesCLI class: ``cli`` for installed/imported entry
+            # points, ``__main__`` for ``python cli.py``/Fire.  Assign that
+            # owner module explicitly; a bare ``global`` here would create/update
+            # a shadow variable in this mixin module and leave cleanup blind to
+            # the live agent.
+            _cli_mod = sys.modules.get(self.__class__.__module__)
+            if _cli_mod is None or not hasattr(_cli_mod, "_active_agent_ref"):
+                _cli_mod = sys.modules.get("cli")
+            if _cli_mod is not None:
+                setattr(_cli_mod, "_active_agent_ref", self.agent)
             # Route agent status output through prompt_toolkit so ANSI escape
             # sequences aren't garbled by patch_stdout's StdoutProxy (#2262).
             self.agent._print_fn = _cprint
diff --git a/run_agent.py b/run_agent.py
index e81bf3b93e7ee..d3def35572fe0 100644
--- a/run_agent.py
+++ b/run_agent.py
@@ -2999,10 +2999,11 @@ def _sync_external_memory_for_turn(
                 final_response,
                 **sync_kwargs,
             )
-            self._memory_manager.queue_prefetch_all(
-                original_user_message,
-                session_id=self.session_id or "",
-            )
+            if not getattr(self, "_suppress_post_turn_prefetch", False):
+                self._memory_manager.queue_prefetch_all(
+                    original_user_message,
+                    session_id=self.session_id or "",
+                )
         except Exception:
             pass
 
diff --git a/tests/agent/test_memory_post_turn_prefetch.py b/tests/agent/test_memory_post_turn_prefetch.py
new file mode 100644
index 0000000000000..b10e4431cee1a
--- /dev/null
+++ b/tests/agent/test_memory_post_turn_prefetch.py
@@ -0,0 +1,59 @@
+"""Regression tests for suppressing useless post-turn memory prefetch.
+
+Single-query CLI sessions end immediately after the response.  Queueing a
+"next turn" prefetch in that path starts Honcho background threads that have no
+consumer and can race interpreter shutdown.
+"""
+
+from __future__ import annotations
+
+from run_agent import AIAgent
+
+
+class _RecordingMemoryManager:
+    def __init__(self):
+        self.sync_calls = []
+        self.prefetch_calls = []
+
+    def sync_all(self, user_content, assistant_content, **kwargs):
+        self.sync_calls.append((user_content, assistant_content, kwargs))
+
+    def queue_prefetch_all(self, query, **kwargs):
+        self.prefetch_calls.append((query, kwargs))
+
+
+def _agent_with_memory_manager(suppress_prefetch: bool = False):
+    agent = AIAgent.__new__(AIAgent)
+    setattr(agent, "_memory_manager", _RecordingMemoryManager())
+    setattr(agent, "session_id", "sess-1")
+    setattr(agent, "_suppress_post_turn_prefetch", suppress_prefetch)
+    return agent
+
+
+def test_external_memory_sync_can_suppress_next_turn_prefetch():
+    agent = _agent_with_memory_manager(suppress_prefetch=True)
+
+    agent._sync_external_memory_for_turn(
+        original_user_message="hello",
+        final_response="world",
+        interrupted=False,
+        messages=[{"role": "user", "content": "hello"}],
+    )
+
+    mm = getattr(agent, "_memory_manager")
+    assert len(mm.sync_calls) == 1
+    assert mm.prefetch_calls == []
+
+
+def test_external_memory_sync_prefetches_by_default():
+    agent = _agent_with_memory_manager(suppress_prefetch=False)
+
+    agent._sync_external_memory_for_turn(
+        original_user_message="hello",
+        final_response="world",
+        interrupted=False,
+    )
+
+    mm = getattr(agent, "_memory_manager")
+    assert len(mm.sync_calls) == 1
+    assert mm.prefetch_calls == [("hello", {"session_id": "sess-1"})]
diff --git a/tests/cli/test_cli_active_agent_ref.py b/tests/cli/test_cli_active_agent_ref.py
new file mode 100644
index 0000000000000..e8b466262ac57
--- /dev/null
+++ b/tests/cli/test_cli_active_agent_ref.py
@@ -0,0 +1,146 @@
+"""Regression tests for CLI active-agent cleanup wiring.
+
+The agent construction mixin lives in ``hermes_cli.cli_agent_setup_mixin`` but
+process cleanup reads ``_active_agent_ref`` from the concrete CLI owner module.
+A module-local assignment in the mixin leaves cleanup blind to the live agent, so
+memory-provider background threads can survive until interpreter shutdown.
+"""
+
+from __future__ import annotations
+
+import sys
+import types
+
+import cli as cli_mod
+from hermes_cli.cli_agent_setup_mixin import CLIAgentSetupMixin
+
+
+class _FakeAgent:
+    def __init__(self, **kwargs):
+        self.kwargs = kwargs
+        self.session_id = kwargs.get("session_id")
+        self._print_fn = None
+
+
+class _DummyCLI(CLIAgentSetupMixin):
+    def __init__(self):
+        self.agent = None
+        self.api_key = "test-key"
+        self.base_url = "https://example.invalid"
+        self.provider = "openai"
+        self.api_mode = "chat"
+        self.acp_command = None
+        self.acp_args = []
+        self.model = "test-model"
+        self.max_tokens = None
+        self.max_turns = 3
+        self.enabled_toolsets = []
+        self.disabled_toolsets = []
+        self.verbose = False
+        self.tool_progress_mode = "off"
+        self.system_prompt = None
+        self.prefill_messages = []
+        self.reasoning_config = None
+        self.service_tier = None
+        self._providers_only = None
+        self._providers_ignore = None
+        self._providers_order = None
+        self._provider_sort = None
+        self._provider_require_params = False
+        self._provider_data_collection = None
+        self._openrouter_min_coding_score = None
+        self.session_id = "test-session"
+        self._session_db = object()
+        self._fallback_model = None
+        self.checkpoints_enabled = False
+        self.checkpoint_max_snapshots = 0
+        self.checkpoint_max_total_size_mb = 0
+        self.checkpoint_max_file_size_mb = 0
+        self.pass_session_id = False
+        self.ignore_rules = False
+        self.streaming_enabled = False
+        self._inline_diffs_enabled = False
+        self._active_agent_route_signature = None
+        self._pending_title = None
+        self._resumed = False
+        self.conversation_history = []
+        self.requested_provider = None
+        self._explicit_api_key = None
+        self._explicit_base_url = None
+        self._credential_pool = None
+
+    def _install_tool_callbacks(self):
+        pass
+
+    def _ensure_tirith_security(self):
+        pass
+
+    def _ensure_runtime_credentials(self):
+        return True
+
+    def _clarify_callback(self, *args, **kwargs):
+        return ""
+
+    def _current_reasoning_callback(self):
+        return None
+
+    def _on_thinking(self, *args, **kwargs):
+        pass
+
+    def _on_tool_progress(self, *args, **kwargs):
+        pass
+
+    def _on_tool_start(self, *args, **kwargs):
+        pass
+
+    def _on_tool_complete(self, *args, **kwargs):
+        pass
+
+    def _stream_delta(self, *args, **kwargs):
+        pass
+
+    def _on_tool_gen_start(self, *args, **kwargs):
+        pass
+
+    def _on_notice(self, *args, **kwargs):
+        pass
+
+    def _on_notice_clear(self, *args, **kwargs):
+        pass
+
+
+def _patch_agent_startup(monkeypatch):
+    monkeypatch.setattr(cli_mod, "AIAgent", _FakeAgent)
+    monkeypatch.setattr(cli_mod, "_prepare_deferred_agent_startup", lambda: None)
+    monkeypatch.setattr("hermes_cli.mcp_startup.wait_for_mcp_discovery", lambda: None)
+
+
+def test_init_agent_updates_cli_module_active_agent_ref(monkeypatch):
+    _patch_agent_startup(monkeypatch)
+    monkeypatch.setattr(cli_mod, "_active_agent_ref", None)
+
+    dummy = _DummyCLI()
+
+    assert dummy._init_agent() is True
+    assert dummy.agent is not None
+    assert cli_mod._active_agent_ref is dummy.agent
+
+
+def test_init_agent_updates_script_owner_module_active_agent_ref(monkeypatch):
+    _patch_agent_startup(monkeypatch)
+    monkeypatch.setattr(cli_mod, "_active_agent_ref", None)
+    owner = types.ModuleType("fake_cli_script_owner")
+    setattr(owner, "_active_agent_ref", None)
+    monkeypatch.setitem(sys.modules, "fake_cli_script_owner", owner)
+    script_like_cli = type(
+        "ScriptLikeCLI",
+        (_DummyCLI,),
+        {"__module__": "fake_cli_script_owner"},
+    )
+
+    dummy = script_like_cli()
+
+    assert dummy._init_agent() is True
+    assert dummy.agent is not None
+    assert owner._active_agent_ref is dummy.agent
+    assert cli_mod._active_agent_ref is None