diff --git a/raven/cli/_memory_warn.py b/raven/cli/_memory_warn.py
new file mode 100644
index 0000000..fa4cbd2
--- /dev/null
+++ b/raven/cli/_memory_warn.py
@@ -0,0 +1,21 @@
+"""User-visible warning when the memory backend fails to start.
+
+Under the TUI, loguru is redirected to a file, so a start failure (e.g. a
+memory identity that EverOS rejects) silently disables long-term memory with
+nothing on screen. Mirror the embedding-unavailable warning in
+``everos/backend.py`` and print to stderr so every entry point (TUI, agent
+REPL, gateway) sees the same degraded-memory notice, not just the two that
+already surface the traceback via stderr.
+"""
+
+from __future__ import annotations
+
+from rich.console import Console
+
+
+def warn_memory_start_failed(exc: BaseException) -> None:
+    """Print a stderr notice that long-term memory is off this session."""
+    Console(stderr=True).print(
+        "[yellow]Memory backend failed to start; long-term memory is off this session.[/yellow]\n"
+        f"[dim]Run `raven onboard` to reconfigure, or check the log ({type(exc).__name__}).[/dim]"
+    )
diff --git a/raven/cli/agent_commands.py b/raven/cli/agent_commands.py
index 715c64f..6e34300 100644
--- a/raven/cli/agent_commands.py
+++ b/raven/cli/agent_commands.py
@@ -398,10 +398,13 @@ async def run_once():
                 if backend is not None:
                     try:
                         await backend.start()
-                    except Exception:
+                    except Exception as exc:
                         logger.exception(
                             "memory backend start failed; continuing with legacy memory path",
                         )
+                        from raven.cli._memory_warn import warn_memory_start_failed
+
+                        warn_memory_start_failed(exc)
                 try:
                     # Build inside the running loop: Scheduler pins its home loop in
                     # __init__, so build_repl must not run in the sync prologue.
@@ -496,10 +499,13 @@ async def run_interactive():
                 if backend is not None:
                     try:
                         await backend.start()
-                    except Exception:
+                    except Exception as exc:
                         logger.exception(
                             "memory backend start failed; continuing with legacy memory path",
                         )
+                        from raven.cli._memory_warn import warn_memory_start_failed
+
+                        warn_memory_start_failed(exc)
                 # agent_loop.run() is now a lifecycle keep-alive (executor /
                 # debug server / MCP up, then idle); all turns go through the
                 # spine. Gathered on teardown.
diff --git a/raven/cli/gateway_commands.py b/raven/cli/gateway_commands.py
index 9edfccc..ce79351 100644
--- a/raven/cli/gateway_commands.py
+++ b/raven/cli/gateway_commands.py
@@ -368,10 +368,13 @@ async def run():
             if backend is not None:
                 try:
                     await backend.start()
-                except Exception:
+                except Exception as exc:
                     _logger.exception(
                         "memory backend start failed; continuing with legacy memory path",
                     )
+                    from raven.cli._memory_warn import warn_memory_start_failed
+
+                    warn_memory_start_failed(exc)
             try:
                 # Spine assembly for the gateway's host sources (cron submits
                 # through it, replies route to channels via a per-channel outlet).
diff --git a/raven/cli/tui_commands.py b/raven/cli/tui_commands.py
index 3c53367..5cf994d 100644
--- a/raven/cli/tui_commands.py
+++ b/raven/cli/tui_commands.py
@@ -706,12 +706,15 @@ def _agent_loop_factory():
             async def _start_backend() -> None:
                 try:
                     await agent_loop.backend.start()  # type: ignore[union-attr]
-                except Exception:
+                except Exception as exc:
                     from loguru import logger as _logger
 
                     _logger.exception(
                         "tui: memory backend start failed; continuing with degraded memory path",
                     )
+                    from raven.cli._memory_warn import warn_memory_start_failed
+
+                    warn_memory_start_failed(exc)
                 _strip_tty_stream_handlers()
 
             asyncio.create_task(_start_backend())
diff --git a/tests/test_memory_warn.py b/tests/test_memory_warn.py
new file mode 100644
index 0000000..0d1da6d
--- /dev/null
+++ b/tests/test_memory_warn.py
@@ -0,0 +1,19 @@
+"""warn_memory_start_failed must surface on stderr (visible under the TUI)."""
+
+from __future__ import annotations
+
+import pytest
+
+from raven.cli._memory_warn import warn_memory_start_failed
+
+
+def test_warn_memory_start_failed_writes_degraded_notice_to_stderr(
+    capsys: pytest.CaptureFixture[str],
+) -> None:
+    warn_memory_start_failed(ValueError("memory.userId='..' is not accepted"))
+
+    captured = capsys.readouterr()
+    combined = captured.err + captured.out
+
+    assert "long-term memory is off" in combined
+    assert "ValueError" in combined