fix(memory): keep the user's name out of the persona in a turn without memory

changes/552.fixed.md

@@ -0,0 +1,3 @@
+**Connected apps and memory-off chats no longer see your name in Sentient's personality.** The name you gave at setup is
+written into the personality file, so a chat with memory off (a connected app, or an incognito chat) could still reveal
+it. Now that chat reads "the user" there, and keeps the rest of the personality. Normal chats still use your name.

sentient/agent/loop.py

@@ -56,7 +56,7 @@
 from sentient.agent import undo as undo_rules
 from sentient.agent.approvals import ApprovalBroker
 from sentient.agent.chat_rules import WAIT_S, pending_asks, standing_text
-from sentient.agent.prompt import build_system_prompt, frame_api_request, mark_outside
+from sentient.agent.prompt import build_system_prompt, frame_api_request, mark_outside, without_name
 from sentient.agent.sources import AnswerSources
 from sentient.agent.toolselect import ToolSelector, is_local_model
 from sentient.app_names import clean_app_name, framed_app_name
@@ -794,8 +794,11 @@ async def _system_prompt_with_sources(
         sources.add_facts(facts)
         sources.add_insights(insights)
         snapshot = self.workspace.snapshot()
-        if not reads_memory:  # only the persona: no profile, memory file or daily notes
-            snapshot = {k: (v if k == "soul" else "") for k, v in snapshot.items()}
+        if not reads_memory:  # only the persona, without the user's name: no profile, memory file or daily notes (#552)
+            snapshot = {
+                k: (without_name(v, self.config.assistant.user_name) if k == "soul" else "")
+                for k, v in snapshot.items()
+            }
         elif not personal:  # MEMORY.md, daily notes and USER.md's learned facts come from personal memories (#155)
             snapshot = {
                 **snapshot, "memory": "", "today": "", "yesterday": "",

sentient/agent/prompt.py

@@ -27,6 +27,16 @@
 - Keep replies short. Plain language. Match the user's tone.
 """
 
+
+def without_name(text: str, name: str) -> str:
+    """``text`` with the user's ``name`` replaced by "the user" (#552), for a turn that must not reveal who they are:
+    SOUL.md is written with the name in it ("You are Sentient, Maya's assistant..."). Whole-word, any case."""
+    name = (name or "").strip()
+    if not name:
+        return text
+    return re.sub(rf"(?<!\w){re.escape(name)}(?!\w)", "the user", text, flags=re.IGNORECASE)
+
+
 # ---------------------------------------------------------------------- outside content (spotlighting, ADR 0023)
 # Tool results that can carry text someone else wrote (rules.brings_untrusted) are wrapped in markers before the model
 # reads them, and the system prompt says once that marked text is data. The code in the markers comes from a secret

tests/test_memory_off_turns.py

@@ -211,3 +211,36 @@ async def test_a_script_gets_no_memory_tools_from_a_turn_without_memory(app):
     assert res["ok"], res
     assert str(res["result"]["got"]).startswith("refused") and not MEMORY_TOOLS & set(res["result"]["available"])
     assert "Anika" not in json.dumps(res)
+
+
+async def test_a_turn_without_memory_keeps_the_persona_but_not_the_users_name(app):
+    """SOUL.md is written with the name in it at onboarding (#552): without memory the prompt says "the user"."""
+    from sentient.memory.personas import render_persona
+
+    app.workspace.write("soul", render_persona("professional", name="Sentient", user="Maya"))
+    sid = await _private_chat(app)
+    app.fake.replies.append("Hello.")
+    await _turn(app, sid, "what's my name?")
+    private = app.fake.calls[0]["messages"][0]["content"]
+    assert "Maya" not in private
+    assert "the user's chief of staff" in private  # the persona stays
+
+    sid = await app.store.create_session(channel="desktop")
+    app.fake.replies.append("Hello.")
+    await _turn(app, sid, "hi")
+    assert "Maya's chief of staff" in app.fake.calls[-1]["messages"][0]["content"]  # a normal chat still has it
+
+
+async def test_the_v1_api_without_memory_does_not_reveal_the_name_from_soul(app):
+    app.config.openai_api.allow_memory = False
+    assert "Maya" in app.workspace.snapshot()["soul"]  # the default persona names the user
+    prompt = await app.agent.system_prompt("what's my name?", "api")
+    assert "Maya" not in prompt
+
+
+def test_without_name_replaces_whole_words_in_any_case():
+    from sentient.agent.prompt import without_name
+
+    assert without_name("Maya's bot. MAYA and Mayan.", "Maya") == "the user's bot. the user and Mayan."
+    assert without_name("Dr. Maya Rao's bot", "Maya Rao") == "Dr. the user's bot"
+    assert without_name("unchanged", "") == "unchanged"