diff --git a/navi/core/agent.py b/navi/core/agent.py
index bd26560..70f32a9 100644
--- a/navi/core/agent.py
+++ b/navi/core/agent.py
@@ -95,15 +95,6 @@
         tool_schemas = [t.schema() for t in tools]
         llm = self._get_backend(profile.llm_backend)
 
-        # System prompt only goes into context (not display history).
-        # Use role check rather than empty check: backward-compat sessions have
-        # context initialised from messages (which never contain a system message).
-        if not any(m.role == "system" for m in session.context):
-            session.context.insert(0, Message(
-                role="system",
-                content=self._build_system_prompt(profile.system_prompt),
-            ))
-
         mem = await self._memory_msg()
 
         # Expose session_id to tools (e.g. SSH connection pool) via ContextVar
@@ -119,7 +110,7 @@
         for iteration in range(profile.max_iterations):
             log.debug("agent.iteration", session_id=session_id, iteration=iteration)
             response = await llm.complete(
-                self._with_memory(session.context, mem),
+                self._build_context(session.context, profile, mem),
                 tools=tool_schemas if tools else None,
                 temperature=profile.temperature,
                 model=profile.model,
@@ -175,11 +166,10 @@
 
         mem = await self._memory_msg()
 
+        # Sub-agent context: only user/assistant/tool messages — system is injected dynamically.
         context: list[Message] = [
-            Message(role="system", content=self._build_system_prompt(profile.system_prompt))
+            Message(role="user", content=user_message, created_at=datetime.now(timezone.utc))
         ]
-        context.append(Message(role="user", content=user_message,
-                               created_at=datetime.now(timezone.utc)))
 
         # Read the event sink set by the parent run_stream() for this tool call.
         # If None (e.g. called from run(), not run_stream()), events are silently dropped.
@@ -197,7 +187,7 @@
             turn_tool_calls: list[ToolCallRequest] | None = None
 
             async for chunk in llm.stream_complete(
-                self._with_memory(context, mem),
+                self._build_context(context, profile, mem),
                 tools=tool_schemas if tools else None,
                 temperature=profile.temperature,
                 model=profile.model,
@@ -284,15 +274,6 @@
         tool_schemas = [t.schema() for t in tools]
         llm = self._get_backend(profile.llm_backend)
 
-        # System prompt only goes into context (not display history).
-        # Use role check rather than empty check: backward-compat sessions have
-        # context initialised from messages (which never contain a system message).
-        if not any(m.role == "system" for m in session.context):
-            session.context.insert(0, Message(
-                role="system",
-                content=self._build_system_prompt(profile.system_prompt),
-            ))
-
         mem = await self._memory_msg()
 
         # Expose session_id to tools (e.g. SSH connection pool) via ContextVar
@@ -350,7 +331,7 @@
             context_tokens: int | None = None
 
             async for chunk in llm.stream_complete(
-                self._with_memory(session.context, mem),
+                self._build_context(session.context, profile, mem),
                 tools=tool_schemas if tools else None,
                 temperature=profile.temperature,
                 model=profile.model,
@@ -490,13 +471,30 @@
             return None
         return Message(role="system", content=f"## What I remember about the user\n\n{summary}")
 
-    def _with_memory(self, ctx: list[Message], mem: "Message | None") -> list[Message]:
-        """Inject memory message after the first system message without mutating ctx."""
-        if mem is None:
-            return ctx
-        if ctx and ctx[0].role == "system":
-            return [ctx[0], mem] + ctx[1:]
-        return [mem] + ctx
+    def _build_context(
+        self,
+        session_context: list[Message],
+        profile: "AgentProfile",
+        mem: "Message | None",
+    ) -> list[Message]:
+        """Build the full LLM context for one call.
+
+        System prompt is injected fresh from the current profile every time —
+        it is NOT stored in session.context so that profile switches take
+        effect immediately without touching stored history.
+        Memory (if any) is placed right after the system message.
+        Any system messages already in session.context are stripped (migration safety).
+        """
+        system_msg = Message(
+            role="system",
+            content=self._build_system_prompt(profile.system_prompt),
+        )
+        conv = [m for m in session_context if m.role != "system"]
+        result: list[Message] = [system_msg]
+        if mem:
+            result.append(mem)
+        result.extend(conv)
+        return result
 
     def _build_system_prompt(self, profile_prompt: str) -> str:
         persona = settings.navi_persona.strip()