diff --git a/docs/INTERNAL-SUGGESTION-JOBS.md b/docs/INTERNAL-SUGGESTION-JOBS.md
new file mode 100644
index 0000000..d8da341
--- /dev/null
+++ b/docs/INTERNAL-SUGGESTION-JOBS.md
@@ -0,0 +1,20 @@
+# Suggestion job provenance
+
+The default suggestion judge launches `amplifier run --output-format json` with
+these values in a private copy of its child environment:
+
+- `AMPLIFIER_SESSION_VISIBILITY=internal`
+- `AMPLIFIER_SESSION_PURPOSE=memory.suggestion`
+- `AMPLIFIER_SESSION_ORIGIN=agent`
+
+Only this implementation job gets the declaration. The caller's environment,
+provider routing, credentials, CLI arguments, timer scheduling, and retry
+behavior are unchanged. No raw environment values are logged.
+
+The compatible CLI persists visibility and purpose as creation metadata. Unified
+can then keep this history available for diagnostics while excluding it from
+ordinary chat lists and default searches. A generic agent origin does not make
+other standalone CLI conversations internal. End-to-end presentation requires
+both the CLI persistence support and the Unified consumer support; an older CLI
+ignores the visibility declaration. Historical unmarked sessions are not
+rewritten or classified by prompt wording.
diff --git a/src/amplifier_memory/suggest.py b/src/amplifier_memory/suggest.py
index bcbbf54..bf87ca5 100644
--- a/src/amplifier_memory/suggest.py
+++ b/src/amplifier_memory/suggest.py
@@ -1060,12 +1060,22 @@ def default_model_call(prompt: str, *, timeout: float = 300.0, call: Judge | Non
             "`run_suggest` (and point AMPLIFIER_CONTEXT_INTELLIGENCE_BASE_PATH at a fixture)"
         )
     argv = build_argv(prompt, call)
+    # Only this background judge is internal. Preserve routing/credentials in
+    # the child environment without relabelling the caller or other CLI jobs.
+    child_env = dict(os.environ)
+    child_env.update(
+        AMPLIFIER_SESSION_VISIBILITY="internal",
+        AMPLIFIER_SESSION_PURPOSE="memory.suggestion",
+        AMPLIFIER_SESSION_ORIGIN="agent",
+    )
     # The command as run, minus the request itself: a failure names the flags that
     # produced it (a provider id that does not exist on this device is the likely one),
     # never the whole transcript.
     shown = " ".join(argv[:-1])
     try:
-        proc = subprocess.run(argv, capture_output=True, text=True, check=False, timeout=timeout)
+        proc = subprocess.run(
+            argv, capture_output=True, text=True, check=False, timeout=timeout, env=child_env
+        )
     except (OSError, ValueError, subprocess.TimeoutExpired) as exc:
         raise RuntimeError(f"{shown} failed: {type(exc).__name__}: {exc}") from exc
     if proc.returncode != 0:
diff --git a/tests/test_suggest.py b/tests/test_suggest.py
index 7d25311..6c585db 100644
--- a/tests/test_suggest.py
+++ b/tests/test_suggest.py
@@ -1287,3 +1287,34 @@ def test_a_run_over_a_worker_only_night_still_reports(tmp_path: Path, store: Pat
     assert report.status == "ok", "a night with no human sessions is not a degraded night"
     assert "sessions=0 origin_excluded=3" in report.log_line
     assert call.prompts == [], "no session read means no model call"
+
+
+@pytest.mark.parametrize("returncode", [0, 1])
+def test_only_suggestion_child_is_marked_internal(monkeypatch, returncode):
+    """Explicit provenance reaches CLI without mutating this process or retrying."""
+    import os
+
+    seen = []
+    monkeypatch.setenv("AMPLIFIER_SESSION_VISIBILITY", "chat")
+    monkeypatch.setenv("AMPLIFIER_SESSION_PURPOSE", "caller.task")
+    monkeypatch.setenv("AMPLIFIER_SESSION_ORIGIN", "human")
+    monkeypatch.setenv("SYNTHETIC_ROUTING_SETTING", "keep")
+    monkeypatch.delenv("PYTEST_CURRENT_TEST", raising=False)
+    before = dict(os.environ)
+
+    def fake_run(argv, **kwargs):
+        seen.append((argv, kwargs))
+        return subprocess.CompletedProcess(argv, returncode, '{"response":"[]"}', "synthetic failure")
+
+    monkeypatch.setattr(suggest.subprocess, "run", fake_run)
+    if returncode:
+        with pytest.raises(RuntimeError, match="exited 1"):
+            suggest.default_model_call("synthetic")
+    else:
+        assert suggest.default_model_call("synthetic") == "[]"
+    assert len(seen) == 1
+    argv, kwargs = seen[0]
+    assert argv == ["amplifier", "run", "--output-format", "json", "synthetic"]
+    assert kwargs["env"] == {**before, "AMPLIFIER_SESSION_VISIBILITY": "internal",
+        "AMPLIFIER_SESSION_PURPOSE": "memory.suggestion", "AMPLIFIER_SESSION_ORIGIN": "agent"}
+    assert dict(os.environ) == before