diff --git a/src/everos_mcp/server.py b/src/everos_mcp/server.py
index 827c6b2..3c519be 100644
--- a/src/everos_mcp/server.py
+++ b/src/everos_mcp/server.py
@@ -61,6 +61,13 @@
 log = logging.getLogger("everos_mcp")
 
 
+# Extraction names each speaker by sender_name, falling back to sender_id, so
+# without these the stored memories would read "default-user said ..." or, on
+# the HTTP server, quote the user's opaque account id.
+USER_SENDER_NAME = "User"
+ASSISTANT_SENDER_NAME = "Assistant"
+
+
 def _data(body: str) -> str:
     """Fence recalled text: it was written in earlier sessions, possibly from
     untrusted content, so it must read as data. Everything the server itself
@@ -226,12 +233,19 @@ async def add_memory(
     messages = []
     if user_message:
         messages.append(
-            {"sender_id": s.user_id, "role": "user", "timestamp": ts, "content": user_message}
+            {
+                "sender_id": s.user_id,
+                "sender_name": USER_SENDER_NAME,
+                "role": "user",
+                "timestamp": ts,
+                "content": user_message,
+            }
         )
     if assistant_message:
         messages.append(
             {
                 "sender_id": s.assistant_sender_id,
+                "sender_name": ASSISTANT_SENDER_NAME,
                 "role": "assistant",
                 "timestamp": ts + 1,
                 "content": assistant_message,
@@ -434,6 +448,8 @@ async def record_trajectory(messages: list[dict[str, Any]], ctx: Context | None
             "timestamp": ts + i,
             "content": m.get("content") or "",
         }
+        if role != "tool":
+            item["sender_name"] = USER_SENDER_NAME if role == "user" else ASSISTANT_SENDER_NAME
         if m.get("tool_calls"):
             item["tool_calls"] = _normalize_tool_calls(m["tool_calls"])
         if role == "tool":
diff --git a/tests/test_everos_mcp.py b/tests/test_everos_mcp.py
index b6d4373..3deb9a4 100644
--- a/tests/test_everos_mcp.py
+++ b/tests/test_everos_mcp.py
@@ -267,6 +267,17 @@ def test_notices_are_attached_once(tool_client):
     assert "boom" in first and "boom" not in second
 
 
+def test_add_memory_names_both_speakers(tool_client):
+    # Extraction falls back to sender_id when there is no name, which puts
+    # "default-user" or an opaque account id into the stored memory text.
+    seen = tool_client(routes={"add": {"data": {"status": "extracted"}}})
+    run(server.add_memory(user_message="u", assistant_message="a", wait=True))
+    user, assistant = sent(seen[0])["messages"]
+    assert (user["sender_id"], user["sender_name"]) == (SETTINGS.user_id, "User")
+    assert assistant["sender_id"] == SETTINGS.assistant_sender_id
+    assert assistant["sender_name"] == "Assistant"
+
+
 def test_add_memory_refuses_secret_without_network(tool_client):
     seen = tool_client()
     reply = run(call_and_drain(server.add_memory(user_message="key sk-abcdefghijklmnop123")))
@@ -312,6 +323,8 @@ def test_record_trajectory_lands_in_the_conversation_session(tool_client):
     assert payload["session_id"] == SETTINGS.session_id
     assert paths(seen) == ["add", "flush"]
     assert payload["messages"][1]["tool_calls"][0]["function"]["name"] == "t"
+    names = [m.get("sender_name") for m in payload["messages"]]
+    assert names == ["User", "Assistant", None, "Assistant"]
 
 
 def test_record_trajectory_validates_tool_messages(tool_client):