fix: improve runtime memory update diagnostics

agents/planner_node.py

@@ -3,7 +3,9 @@
 from utils.language import (
     contains_cyrillic,
 )
-
+from settings.config_loader import (
+    config,
+)
 
 class PlannerNode(BaseNode):
 
@@ -13,6 +15,18 @@ async def run(
             context,
     ):
 
+        if not getattr(config, "TRANSLATION_ENABLED", False):
+            state.current_plan = [
+                "brain",
+                "validator",
+            ]
+
+            state.translate_input = False
+            state.translated_input = state.user_input
+
+
+            return
+
         state.translate_input = contains_cyrillic(
             state.user_input
         )

agents/translation_node.py

@@ -24,22 +24,18 @@ async def run(
         state.iteration += 1
 
         translated = await translate(
-            client=context.clients[
-                "translator"
-            ],
+            context=context,
             text=state.user_input,
             source_language="Russian",
             target_language="English",
         )
 
-        translated_text = translated[
-            "content"
-        ]
-
-        usage = translated.get(
-            "usage",
-            {},
-        )
+        if isinstance(translated, str):
+            translated_text = translated
+            usage = {}
+        else:
+            translated_text = translated["content"]
+            usage = translated.get("usage", {})
 
         await context.logger.log_translation(
             translated_text

clients/translation_client.py

@@ -39,12 +39,14 @@ def build_translation_system_prompt(
 
 async def translate(
     *,
-    client,
+    context,
     text: str,
     source_language: str,
     target_language: str,
 ):
-
+    client=context.clients[
+        "translator"
+    ]
     stage = (
         f"{source_language}"
         f"_to_"
@@ -82,23 +84,15 @@ async def translate(
         )
 
         if content:
-
             return {
                 "content": content,
-                "usage": (
-                    result.get(
-                        "usage",
-                        {},
-                    )
-                ),
+                "usage": result.get("usage", {}),
             }
 
-        return (
-            ResponseExtractor
-            .extract_reasoning_text(
-                result
-            )
-        )
+        return {
+            "content": text,
+            "usage": result.get("usage", {}),
+        }
 
     except asyncio.CancelledError:
         raise

config.example.py

@@ -1,6 +1,7 @@
 # Copy this file to config.py and adjust values for your local nodes.
 
 USE_SERVICE_AS_BRAIN = False
+TRANSLATION_ENABLED = False
 
 CHAT_ENDPOINT = "/v1/chat/completions"
 MODELS_ENDPOINT = "/v1/models"
@@ -21,7 +22,7 @@
 
 BRAIN_TEMPERATURE = 0.7
 
-BRAIN_MAX_TOKENS = 2048
+BRAIN_MAX_TOKENS = 8192
 
 # ---------------------------------------------------------
 # SERVICE MODEL
@@ -37,7 +38,7 @@
 
 SERVICE_TEMPERATURE = 0.1
 
-SERVICE_MAX_TOKENS = 1024
+SERVICE_MAX_TOKENS = 4096
 
 # ---------------------------------------------------------
 # SEARCH

memory/message_memory.py

@@ -82,7 +82,12 @@ def build_runtime_memory_system_prompt() -> str:
         "Avoid writing about JIN's role unless the role itself changed or matters. "
         "Describe assistant actions neutrally instead.\n"
         "Keep memory actionable: write what helps the next answer, not a recap of "
-        "what happened.\n"
+        "what happened. \n"
+        "If there are unresolved pending choices or open references "
+        "that remain relevant to the current conversation, "
+        "you may naturally remind the user about them.\n"
+        "Do not interrupt a clearly established new topic. "
+        "Use reminders sparingly and only when they add value.\n"
         "Do not merge unrelated facts into one sentence. Prefer separate lines "
         "over broad phrasing like 'Topic established: X, specifically Y'.\n"
         "Finish every bullet line completely. Never leave a line mid-phrase.\n"
@@ -406,8 +411,10 @@ async def summarize_runtime_memory(
                     None,
                 ),
                 "[MEMORY] runtime memory update skipped",
-                details=(
-                    "Summarizer returned an incomplete memory update."
+                details=build_memory_update_skip_details(
+                    reason="Summarizer returned an incomplete memory update.",
+                    previous_memory=current_memory,
+                    candidate_memory=updated_memory,
                 ),
             )
 
@@ -533,14 +540,15 @@ async def summarize_runtime_memory_pending_turns(
             response
         )
 
-        if (
-                is_runtime_memory_response_truncated(
-                    response
-                )
-                or looks_like_incomplete_runtime_memory(
-            updated_memory
-        )
-        ):
+        skip_reason = None
+
+        if is_runtime_memory_response_truncated(response):
+            skip_reason = "Summarizer response was truncated by max_tokens."
+
+        elif looks_like_incomplete_runtime_memory(updated_memory):
+            skip_reason = "Summarizer returned text that looks structurally incomplete."
+
+        if skip_reason:
             await safe_call(
                 getattr(
                     getattr(
@@ -552,8 +560,10 @@ async def summarize_runtime_memory_pending_turns(
                     None,
                 ),
                 "[MEMORY] runtime memory update skipped",
-                details=(
-                    "Summarizer returned an incomplete memory update."
+                details=build_memory_update_skip_details(
+                    reason="Summarizer returned an incomplete memory update.",
+                    previous_memory=initial_memory,
+                    candidate_memory=updated_memory,
                 ),
             )
 
@@ -756,4 +766,21 @@ async def cancel_runtime_memory_update(
     ):
         await task
 
-    context.runtime_memory_update_task = None
+    context.runtime_memory_update_task = None
+
+def build_memory_update_skip_details(
+        *,
+        reason: str,
+        previous_memory: str,
+        candidate_memory: str,
+) -> str:
+
+    return (
+        f"{reason}\n\n"
+        "Previous memory:\n"
+        "----------------\n"
+        f"{previous_memory.strip() or DEFAULT_RUNTIME_MEMORY}\n\n"
+        "Candidate memory:\n"
+        "-----------------\n"
+        f"{candidate_memory.strip() or '<empty>'}"
+    )

tests/test_agent_routing.py

@@ -16,26 +16,29 @@
 class AgentRoutingTests(
     unittest.IsolatedAsyncioTestCase
 ):
-
-    async def test_cyrillic_input_routes_through_translation(self):
+    async def test_cyrillic_input_routes_directly_to_brain(self):
 
         state = AgentState(
-            user_input="\u043f\u0440\u0438\u0432\u0435\u0442"
+            user_input="привет"
         )
 
         await PlannerNode().run(
             state,
             context=None,
         )
 
-        self.assertTrue(
+        self.assertFalse(
             state.translate_input
         )
 
+        self.assertEqual(
+            state.translated_input,
+            "привет",
+        )
+
         self.assertEqual(
             state.current_plan,
             [
-                "translator",
                 "brain",
                 "validator",
             ],

websocket.py

@@ -255,10 +255,16 @@ async def process_message(
             "type": "agent_runtime_end",
         })
 
+        assistant_message = (
+                state.final_answer
+                or state.brain_response
+                or context.runtime_turn_assistant_response
+        )
+
         schedule_runtime_memory_update(
             context=context,
             user_message=user_text,
-            assistant_message=state.brain_response,
+            assistant_message=assistant_message,
         )
 
     except asyncio.CancelledError: