diff --git a/hermes_cli/web_server.py b/hermes_cli/web_server.py
index 8c33a383e5..6298274d78 100644
--- a/hermes_cli/web_server.py
+++ b/hermes_cli/web_server.py
@@ -761,20 +761,32 @@ async def search_sessions(q: str = "", limit: int = 20):
                 else:
                     terms.append(token + "*")
             prefix_query = " ".join(terms)
-            matches = db.search_messages(query=prefix_query, limit=limit)
-            # Group by session_id — return unique sessions with their best snippet
+            # search_messages() limits raw message hits, but this endpoint returns
+            # unique sessions. Overfetch in batches until we either collect the
+            # requested number of unique sessions or exhaust the match list.
             seen: dict = {}
-            for m in matches:
-                sid = m["session_id"]
-                if sid not in seen:
-                    seen[sid] = {
-                        "session_id": sid,
-                        "snippet": m.get("snippet", ""),
-                        "role": m.get("role"),
-                        "source": m.get("source"),
-                        "model": m.get("model"),
-                        "session_started": m.get("session_started"),
-                    }
+            offset = 0
+            batch_size = max(limit, 20)
+            while len(seen) < limit:
+                matches = db.search_messages(query=prefix_query, limit=batch_size, offset=offset)
+                if not matches:
+                    break
+                for m in matches:
+                    sid = m["session_id"]
+                    if sid not in seen:
+                        seen[sid] = {
+                            "session_id": sid,
+                            "snippet": m.get("snippet", ""),
+                            "role": m.get("role"),
+                            "source": m.get("source"),
+                            "model": m.get("model"),
+                            "session_started": m.get("session_started"),
+                        }
+                        if len(seen) >= limit:
+                            break
+                if len(matches) < batch_size:
+                    break
+                offset += batch_size
             return {"results": list(seen.values())}
         finally:
             db.close()
diff --git a/hermes_state.py b/hermes_state.py
index ed95d25f45..721b375a7a 100644
--- a/hermes_state.py
+++ b/hermes_state.py
@@ -1292,7 +1292,7 @@ class SessionDB:
             JOIN messages m ON m.id = messages_fts.rowid
             JOIN sessions s ON s.id = m.session_id
             WHERE {where_sql}
-            ORDER BY rank
+            ORDER BY rank, m.id
             LIMIT ? OFFSET ?
         """
 
diff --git a/tests/hermes_cli/test_web_server.py b/tests/hermes_cli/test_web_server.py
index e83f5bdeb3..3f9c964eb1 100644
--- a/tests/hermes_cli/test_web_server.py
+++ b/tests/hermes_cli/test_web_server.py
@@ -817,6 +817,25 @@ class TestNewEndpoints:
         except Exception:
             pass
 
+    def test_session_search_returns_limit_unique_sessions(self):
+        from hermes_state import SessionDB
+
+        db = SessionDB()
+        try:
+            db.create_session(session_id="session-a", source="cli", model="anthropic/claude-sonnet-4")
+            db.create_session(session_id="session-b", source="cli", model="anthropic/claude-sonnet-4")
+            for _ in range(25):
+                db.append_message("session-a", role="user", content="needle exact phrase")
+            db.append_message("session-b", role="user", content="needle exact phrase")
+        finally:
+            db.close()
+
+        resp = self.client.get('/api/sessions/search?q="needle exact phrase"&limit=2')
+
+        assert resp.status_code == 200
+        data = resp.json()
+        assert [result["session_id"] for result in data["results"]] == ["session-a", "session-b"]
+
 
 # ---------------------------------------------------------------------------
 # Model context length: normalize/denormalize + /api/model/info