feat(memory): automatische Erinnerungs-Extraktion aus Gespraechen

- app/core/memory_extractor.py: LLM destilliert nach je N Turns dauerhafte
  Fakten/Vorlieben aus dem Verlauf, dedupliziert gegen vorhandene Erinnerungen
  und legt sie ab - best-effort, nicht-blockierend (Hintergrund-Task), eigener
  Extraktions-Prompt (JSON, Reasoning aus), Cap-Begrenzung
- Trigger in /api/chat und /ws/voice nach dem Persistieren des Turns
- Konfig: MEMORY_EXTRACTION_ENABLED/_EVERY_N_TURNS/_MAX/_PROVIDER
- Tests: Extraktion, Dedup, kaputtes JSON, Cap, leeres Gespraech, Scheduling
- Doku: README + Architektur-Roadmap (Punkt 3 erledigt)

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

tests/test_memory_extraction.py

@@ -0,0 +1,135 @@
+import asyncio
+
+import pytest
+
+import app.dependencies as deps
+from app.config import settings
+from app.core import memory_extractor as me
+
+
+class StubLLM:
+    def __init__(self, raw):
+        self.raw = raw
+        self.calls = 0
+
+    async def complete(self, text, history=None, session_id=None):
+        self.calls += 1
+        return self.raw
+
+
+@pytest.fixture(autouse=True)
+def _clear_turn_counts():
+    me._turn_counts.clear()
+    yield
+    me._turn_counts.clear()
+
+
+def _seed_conversation(store, session_id="conv", user_id="anonymous"):
+    store.append_message(session_id, user_id, "user", "Ich heisse Anna und wohne in Kiel.")
+    store.append_message(session_id, user_id, "assistant", "Schoen, Anna!")
+
+
+def test_parse_facts_variants():
+    assert me.parse_facts('["heisst Anna", "wohnt in Kiel"]') == ["heisst Anna", "wohnt in Kiel"]
+    # umschlossen von Geschwafel/Markdown
+    assert me.parse_facts('Hier:\n```json\n["x"]\n```') == ["x"]
+    assert me.parse_facts("[]") == []
+    assert me.parse_facts("kein json") == []
+    assert me.parse_facts("") == []
+    # Nicht-Strings werden ignoriert
+    assert me.parse_facts('["ok", 5, null, "  "]') == ["ok"]
+
+
+def test_extracts_and_stores_new_facts(monkeypatch):
+    store = deps.get_store()
+    _seed_conversation(store)
+    monkeypatch.setattr(me, "_build_extractor_llm",
+                        lambda cfg: StubLLM('["heisst Anna", "wohnt in Kiel"]'))
+
+    added = asyncio.run(me.extract_and_store(store, "anonymous", "conv", settings))
+
+    assert added == 2
+    contents = [m.content for m in store.get_memories("anonymous")]
+    assert contents == ["heisst Anna", "wohnt in Kiel"]
+
+
+def test_dedup_skips_known(monkeypatch):
+    store = deps.get_store()
+    _seed_conversation(store)
+    store.add_memory("anonymous", "heisst Anna")
+    # LLM liefert einen bekannten (anders gross-/kleingeschrieben) + einen neuen Fakt.
+    monkeypatch.setattr(me, "_build_extractor_llm",
+                        lambda cfg: StubLLM('["Heisst Anna", "wohnt in Kiel"]'))
+
+    added = asyncio.run(me.extract_and_store(store, "anonymous", "conv", settings))
+
+    assert added == 1
+    contents = [m.content for m in store.get_memories("anonymous")]
+    assert contents == ["heisst Anna", "wohnt in Kiel"]
+
+
+def test_malformed_output_no_crash(monkeypatch):
+    store = deps.get_store()
+    _seed_conversation(store)
+    monkeypatch.setattr(me, "_build_extractor_llm",
+                        lambda cfg: StubLLM("Tut mir leid, kein JSON hier."))
+
+    added = asyncio.run(me.extract_and_store(store, "anonymous", "conv", settings))
+
+    assert added == 0
+    assert store.get_memories("anonymous") == []
+
+
+def test_cap_respected(monkeypatch):
+    store = deps.get_store()
+    _seed_conversation(store)
+    monkeypatch.setattr(settings, "memory_extraction_max", 1)
+    monkeypatch.setattr(me, "_build_extractor_llm",
+                        lambda cfg: StubLLM('["fakt a", "fakt b", "fakt c"]'))
+
+    added = asyncio.run(me.extract_and_store(store, "anonymous", "conv", settings))
+
+    assert added == 1
+    assert len(store.get_memories("anonymous")) == 1
+
+
+def test_empty_conversation_skips_llm(monkeypatch):
+    store = deps.get_store()
+    called = StubLLM("[]")
+    monkeypatch.setattr(me, "_build_extractor_llm", lambda cfg: called)
+
+    added = asyncio.run(me.extract_and_store(store, "anonymous", "leer", settings))
+
+    assert added == 0
+    assert called.calls == 0  # ohne Gespraech kein LLM-Aufruf
+
+
+def test_schedule_only_every_n_turns(monkeypatch):
+    store = deps.get_store()
+    monkeypatch.setattr(settings, "memory_extraction_enabled", True)
+    monkeypatch.setattr(settings, "memory_extraction_every_n_turns", 3)
+
+    async def run():
+        # Session "s1" hat keine Nachrichten -> der geplante Task endet sofort (kein LLM).
+        results = [me.maybe_schedule_extraction(store, "anonymous", "s1") for _ in range(3)]
+        for task in results:
+            if task is not None:
+                await task
+        return results
+
+    tasks = asyncio.run(run())
+    # nur der 3. Aufruf plant einen Task
+    assert tasks[0] is None and tasks[1] is None
+    assert tasks[2] is not None
+
+
+def test_schedule_disabled_is_noop(monkeypatch):
+    store = deps.get_store()
+    monkeypatch.setattr(settings, "memory_extraction_enabled", False)
+    assert me.maybe_schedule_extraction(store, "anonymous", "s1") is None
+
+
+def test_schedule_without_session_is_noop(monkeypatch):
+    store = deps.get_store()
+    monkeypatch.setattr(settings, "memory_extraction_enabled", True)
+    assert me.maybe_schedule_extraction(store, "anonymous", None) is None