fix(retain): preserve normalized experience fact types

hindsight-api-slim/hindsight_api/engine/retain/fact_extraction.py

@@ -1967,7 +1967,7 @@ def get_value(field_name):
         for fact_from_llm in chunk_facts:
             extracted_fact = ExtractedFactType(
                 fact_text=fact_from_llm.fact,
-                fact_type="experience" if fact_from_llm.fact_type == "assistant" else "world",
+                fact_type=fact_from_llm.fact_type,
                 entities=[e.text for e in (fact_from_llm.entities or [])],
                 occurred_start=_parse_datetime(fact_from_llm.occurred_start) if fact_from_llm.occurred_start else None,
                 occurred_end=_parse_datetime(fact_from_llm.occurred_end) if fact_from_llm.occurred_end else None,
@@ -2152,7 +2152,7 @@ async def extract_facts_from_contents(
                     # mentioned_at is always the event_date (when the conversation/document occurred)
                     extracted_fact = ExtractedFactType(
                         fact_text=fact_from_llm.fact,
-                        fact_type="experience" if fact_from_llm.fact_type == "assistant" else "world",
+                        fact_type=fact_from_llm.fact_type,
                         entities=[e.text for e in (fact_from_llm.entities or [])],
                         # occurred_start/end: from LLM only, leave None if not provided
                         occurred_start=_parse_datetime(fact_from_llm.occurred_start)

hindsight-api-slim/tests/test_fact_extraction_experience_regression.py

@@ -0,0 +1,116 @@
+from datetime import datetime, timezone
+from unittest.mock import AsyncMock, MagicMock, patch
+
+import pytest
+
+from hindsight_api.config import HindsightConfig, _get_raw_config
+from hindsight_api.engine.response_models import TokenUsage
+from hindsight_api.engine.retain.fact_extraction import (
+    Fact,
+    RetainContent,
+    extract_facts_from_contents,
+    extract_facts_from_contents_batch_api,
+)
+
+
+@pytest.mark.asyncio
+async def test_extract_facts_from_contents_preserves_experience_fact_type():
+    """
+    The extractor already normalizes raw assistant facts to "experience".
+    The final ExtractedFactType conversion must preserve that normalized type.
+    """
+    contents = [
+        RetainContent(
+            content="I fixed the failing tests after discovering they mocked the wrong interface.",
+            event_date=datetime(2026, 4, 1, tzinfo=timezone.utc),
+            context="assistant work log",
+        )
+    ]
+    extracted_fact = Fact(
+        fact="Assistant fixed the failing tests after discovering they mocked the wrong interface.",
+        fact_type="experience",
+    )
+
+    with patch(
+        "hindsight_api.engine.retain.fact_extraction.extract_facts_from_text",
+        new=AsyncMock(return_value=([extracted_fact], [(contents[0].content, 1)], TokenUsage())),
+    ):
+        facts, chunks, usage = await extract_facts_from_contents(
+            contents=contents,
+            llm_config=None,
+            agent_name="TestAgent",
+            config=_get_raw_config(),
+        )
+
+    assert len(facts) == 1
+    assert len(chunks) == 1
+    assert usage.total_tokens == 0
+    assert facts[0].fact_type == "experience"
+
+
+@pytest.mark.asyncio
+async def test_extract_facts_from_contents_batch_api_preserves_experience_fact_type():
+    """
+    Batch extraction normalizes raw "assistant" facts before conversion.
+    The batch conversion layer must not remap the normalized type to "world".
+    """
+    llm_config = MagicMock()
+    llm_config.provider = "openai"
+    llm_config.model = "gpt-4.1-mini"
+    llm_config._provider_impl = AsyncMock()
+    llm_config._provider_impl.submit_batch = AsyncMock(
+        return_value={
+            "batch_id": "batch_experience_123",
+            "status": "validating",
+            "request_counts": {"total": 1, "completed": 0, "failed": 0},
+        }
+    )
+    llm_config._provider_impl.get_batch_status = AsyncMock(
+        return_value={"status": "completed", "request_counts": {"total": 1, "completed": 1, "failed": 0}}
+    )
+    llm_config._provider_impl.retrieve_batch_results = AsyncMock(
+        return_value=[
+            {
+                "custom_id": "chunk_0",
+                "response": {
+                    "body": {
+                        "choices": [
+                            {
+                                "message": {
+                                    "content": (
+                                        '{"facts":[{"what":"Assistant fixed the failing tests after discovering they '
+                                        'mocked the wrong interface","when":"2026-04-01","where":"N/A",'
+                                        '"who":"assistant","why":"Tests were failing due to incorrect mocking",'
+                                        '"fact_type":"assistant","fact_kind":"event","occurred_start":'
+                                        '"2026-04-01T00:00:00+00:00","occurred_end":"2026-04-01T00:00:00+00:00"}]}'
+                                    )
+                                }
+                            }
+                        ],
+                        "usage": {"prompt_tokens": 10, "completion_tokens": 5, "total_tokens": 15},
+                    }
+                },
+            }
+        ]
+    )
+
+    facts, chunks, usage = await extract_facts_from_contents_batch_api(
+        contents=[
+            RetainContent(
+                content="I fixed the failing tests after discovering they mocked the wrong interface.",
+                event_date=datetime(2026, 4, 1, tzinfo=timezone.utc),
+                context="assistant work log",
+            )
+        ],
+        llm_config=llm_config,
+        agent_name="TestAgent",
+        config=HindsightConfig.from_env(),
+        pool=None,
+        operation_id=None,
+        schema=None,
+    )
+
+    assert len(facts) == 1
+    assert len(chunks) == 1
+    assert usage.total_tokens == 15
+    assert facts[0].fact_type == "experience"