feat(memory): structured reflection + correction detection in MemoryMiddleware (#1620) (#1668)

* feat(memory): add structured reflection and correction detection

* fix(memory): align sourceError schema and prompt guidance

---------

Co-authored-by: Willem Jiang <willem.jiang@gmail.com>

backend/tests/test_memory_prompt_injection.py

@@ -119,3 +119,38 @@ def test_format_memory_skips_non_string_content_facts() -> None:
     # The formatted line for a list content would be "- [knowledge | 0.85] ['list']".
     assert "| 0.85]" not in result
     assert "Valid fact" in result
+
+
+def test_format_memory_renders_correction_source_error() -> None:
+    memory_data = {
+        "facts": [
+            {
+                "content": "Use make dev for local development.",
+                "category": "correction",
+                "confidence": 0.95,
+                "sourceError": "The agent previously suggested npm start.",
+            }
+        ]
+    }
+
+    result = format_memory_for_injection(memory_data, max_tokens=2000)
+
+    assert "Use make dev for local development." in result
+    assert "avoid: The agent previously suggested npm start." in result
+
+
+def test_format_memory_renders_correction_without_source_error_normally() -> None:
+    memory_data = {
+        "facts": [
+            {
+                "content": "Use make dev for local development.",
+                "category": "correction",
+                "confidence": 0.95,
+            }
+        ]
+    }
+
+    result = format_memory_for_injection(memory_data, max_tokens=2000)
+
+    assert "Use make dev for local development." in result
+    assert "avoid:" not in result

backend/tests/test_memory_queue.py

@@ -0,0 +1,50 @@
+from unittest.mock import MagicMock, patch
+
+from deerflow.agents.memory.queue import ConversationContext, MemoryUpdateQueue
+from deerflow.config.memory_config import MemoryConfig
+
+
+def _memory_config(**overrides: object) -> MemoryConfig:
+    config = MemoryConfig()
+    for key, value in overrides.items():
+        setattr(config, key, value)
+    return config
+
+
+def test_queue_add_preserves_existing_correction_flag_for_same_thread() -> None:
+    queue = MemoryUpdateQueue()
+
+    with (
+        patch("deerflow.agents.memory.queue.get_memory_config", return_value=_memory_config(enabled=True)),
+        patch.object(queue, "_reset_timer"),
+    ):
+        queue.add(thread_id="thread-1", messages=["first"], correction_detected=True)
+        queue.add(thread_id="thread-1", messages=["second"], correction_detected=False)
+
+    assert len(queue._queue) == 1
+    assert queue._queue[0].messages == ["second"]
+    assert queue._queue[0].correction_detected is True
+
+
+def test_process_queue_forwards_correction_flag_to_updater() -> None:
+    queue = MemoryUpdateQueue()
+    queue._queue = [
+        ConversationContext(
+            thread_id="thread-1",
+            messages=["conversation"],
+            agent_name="lead_agent",
+            correction_detected=True,
+        )
+    ]
+    mock_updater = MagicMock()
+    mock_updater.update_memory.return_value = True
+
+    with patch("deerflow.agents.memory.updater.MemoryUpdater", return_value=mock_updater):
+        queue._process_queue()
+
+    mock_updater.update_memory.assert_called_once_with(
+        messages=["conversation"],
+        thread_id="thread-1",
+        agent_name="lead_agent",
+        correction_detected=True,
+    )

backend/tests/test_memory_router.py

@@ -72,6 +72,56 @@ def test_import_memory_route_returns_imported_memory() -> None:
     assert response.json()["facts"] == imported_memory["facts"]
 
 
+def test_export_memory_route_preserves_source_error() -> None:
+    app = FastAPI()
+    app.include_router(memory.router)
+    exported_memory = _sample_memory(
+        facts=[
+            {
+                "id": "fact_correction",
+                "content": "Use make dev for local development.",
+                "category": "correction",
+                "confidence": 0.95,
+                "createdAt": "2026-03-20T00:00:00Z",
+                "source": "thread-1",
+                "sourceError": "The agent previously suggested npm start.",
+            }
+        ]
+    )
+
+    with patch("app.gateway.routers.memory.get_memory_data", return_value=exported_memory):
+        with TestClient(app) as client:
+            response = client.get("/api/memory/export")
+
+    assert response.status_code == 200
+    assert response.json()["facts"][0]["sourceError"] == "The agent previously suggested npm start."
+
+
+def test_import_memory_route_preserves_source_error() -> None:
+    app = FastAPI()
+    app.include_router(memory.router)
+    imported_memory = _sample_memory(
+        facts=[
+            {
+                "id": "fact_correction",
+                "content": "Use make dev for local development.",
+                "category": "correction",
+                "confidence": 0.95,
+                "createdAt": "2026-03-20T00:00:00Z",
+                "source": "thread-1",
+                "sourceError": "The agent previously suggested npm start.",
+            }
+        ]
+    )
+
+    with patch("app.gateway.routers.memory.import_memory_data", return_value=imported_memory):
+        with TestClient(app) as client:
+            response = client.post("/api/memory/import", json=imported_memory)
+
+    assert response.status_code == 200
+    assert response.json()["facts"][0]["sourceError"] == "The agent previously suggested npm start."
+
+
 def test_clear_memory_route_returns_cleared_memory() -> None:
     app = FastAPI()
     app.include_router(memory.router)

backend/tests/test_memory_updater.py

@@ -146,6 +146,53 @@ def test_apply_updates_preserves_threshold_and_max_facts_trimming() -> None:
     assert result["facts"][1]["source"] == "thread-9"
 
 
+def test_apply_updates_preserves_source_error() -> None:
+    updater = MemoryUpdater()
+    current_memory = _make_memory()
+    update_data = {
+        "newFacts": [
+            {
+                "content": "Use make dev for local development.",
+                "category": "correction",
+                "confidence": 0.95,
+                "sourceError": "The agent previously suggested npm start.",
+            }
+        ]
+    }
+
+    with patch(
+        "deerflow.agents.memory.updater.get_memory_config",
+        return_value=_memory_config(max_facts=100, fact_confidence_threshold=0.7),
+    ):
+        result = updater._apply_updates(current_memory, update_data, thread_id="thread-correction")
+
+    assert result["facts"][0]["sourceError"] == "The agent previously suggested npm start."
+    assert result["facts"][0]["category"] == "correction"
+
+
+def test_apply_updates_ignores_empty_source_error() -> None:
+    updater = MemoryUpdater()
+    current_memory = _make_memory()
+    update_data = {
+        "newFacts": [
+            {
+                "content": "Use make dev for local development.",
+                "category": "correction",
+                "confidence": 0.95,
+                "sourceError": "   ",
+            }
+        ]
+    }
+
+    with patch(
+        "deerflow.agents.memory.updater.get_memory_config",
+        return_value=_memory_config(max_facts=100, fact_confidence_threshold=0.7),
+    ):
+        result = updater._apply_updates(current_memory, update_data, thread_id="thread-correction")
+
+    assert "sourceError" not in result["facts"][0]
+
+
 def test_clear_memory_data_resets_all_sections() -> None:
     with patch("deerflow.agents.memory.updater._save_memory_to_file", return_value=True):
         result = clear_memory_data()
@@ -522,3 +569,53 @@ class TestUpdateMemoryStructuredResponse:
             result = updater.update_memory([msg, ai_msg])
 
         assert result is True
+
+    def test_correction_hint_injected_when_detected(self):
+        updater = MemoryUpdater()
+        valid_json = '{"user": {}, "history": {}, "newFacts": [], "factsToRemove": []}'
+        model = self._make_mock_model(valid_json)
+
+        with (
+            patch.object(updater, "_get_model", return_value=model),
+            patch("deerflow.agents.memory.updater.get_memory_config", return_value=_memory_config(enabled=True)),
+            patch("deerflow.agents.memory.updater.get_memory_data", return_value=_make_memory()),
+            patch("deerflow.agents.memory.updater.get_memory_storage", return_value=MagicMock(save=MagicMock(return_value=True))),
+        ):
+            msg = MagicMock()
+            msg.type = "human"
+            msg.content = "No, that's wrong."
+            ai_msg = MagicMock()
+            ai_msg.type = "ai"
+            ai_msg.content = "Understood"
+            ai_msg.tool_calls = []
+
+            result = updater.update_memory([msg, ai_msg], correction_detected=True)
+
+        assert result is True
+        prompt = model.invoke.call_args[0][0]
+        assert "Explicit correction signals were detected" in prompt
+
+    def test_correction_hint_empty_when_not_detected(self):
+        updater = MemoryUpdater()
+        valid_json = '{"user": {}, "history": {}, "newFacts": [], "factsToRemove": []}'
+        model = self._make_mock_model(valid_json)
+
+        with (
+            patch.object(updater, "_get_model", return_value=model),
+            patch("deerflow.agents.memory.updater.get_memory_config", return_value=_memory_config(enabled=True)),
+            patch("deerflow.agents.memory.updater.get_memory_data", return_value=_make_memory()),
+            patch("deerflow.agents.memory.updater.get_memory_storage", return_value=MagicMock(save=MagicMock(return_value=True))),
+        ):
+            msg = MagicMock()
+            msg.type = "human"
+            msg.content = "Let's talk about memory."
+            ai_msg = MagicMock()
+            ai_msg.type = "ai"
+            ai_msg.content = "Sure"
+            ai_msg.tool_calls = []
+
+            result = updater.update_memory([msg, ai_msg], correction_detected=False)
+
+        assert result is True
+        prompt = model.invoke.call_args[0][0]
+        assert "Explicit correction signals were detected" not in prompt

backend/tests/test_memory_upload_filtering.py

@@ -10,7 +10,7 @@ persisting in long-term memory:
 from langchain_core.messages import AIMessage, HumanMessage, ToolMessage
 
 from deerflow.agents.memory.updater import _strip_upload_mentions_from_memory
-from deerflow.agents.middlewares.memory_middleware import _filter_messages_for_memory
+from deerflow.agents.middlewares.memory_middleware import _filter_messages_for_memory, detect_correction
 
 # ---------------------------------------------------------------------------
 # Helpers
@@ -134,6 +134,64 @@ class TestFilterMessagesForMemory:
         assert "<uploaded_files>" not in all_content
 
 
+# ===========================================================================
+# detect_correction
+# ===========================================================================
+
+
+class TestDetectCorrection:
+    def test_detects_english_correction_signal(self):
+        msgs = [
+            _human("Please help me run the project."),
+            _ai("Use npm start."),
+            _human("That's wrong, use make dev instead."),
+            _ai("Understood."),
+        ]
+
+        assert detect_correction(msgs) is True
+
+    def test_detects_chinese_correction_signal(self):
+        msgs = [
+            _human("帮我启动项目"),
+            _ai("用 npm start"),
+            _human("不对，改用 make dev"),
+            _ai("明白了"),
+        ]
+
+        assert detect_correction(msgs) is True
+
+    def test_returns_false_without_signal(self):
+        msgs = [
+            _human("Please explain the build setup."),
+            _ai("Here is the build setup."),
+            _human("Thanks, that makes sense."),
+        ]
+
+        assert detect_correction(msgs) is False
+
+    def test_only_checks_recent_messages(self):
+        msgs = [
+            _human("That is wrong, use make dev instead."),
+            _ai("Noted."),
+            _human("Let's discuss tests."),
+            _ai("Sure."),
+            _human("What about linting?"),
+            _ai("Use ruff."),
+            _human("And formatting?"),
+            _ai("Use make format."),
+        ]
+
+        assert detect_correction(msgs) is False
+
+    def test_handles_list_content(self):
+        msgs = [
+            HumanMessage(content=["That is wrong,", {"type": "text", "text": "use make dev instead."}]),
+            _ai("Updated."),
+        ]
+
+        assert detect_correction(msgs) is True
+
+
 # ===========================================================================
 # _strip_upload_mentions_from_memory
 # ===========================================================================