WIP: memory system improvements (in progress)

Interview/chat prompt layers, reply planner, style profiles, memory
injection, interview meta store, and related tests. Work not finished.

Made-with: Cursor

api/app/agents/chat/orchestrator.py

@@ -52,14 +52,13 @@ _UNAUTH_TURN = AgentChatTurn(
 )
 
 
-async def _fetch_interview_memory_evidence(
+async def _fetch_interview_memory_bundle(
     db: AsyncSession,
     user_id: str,
     user_message: str,
-) -> tuple[str, dict | None]:
-    """按本轮用户话检索记忆：格式化短文本 + 可入库 trace（稳定 id）。"""
+) -> tuple[dict | None, object | None]:
+    """检索记忆 bundle（原始结构）；是否进主 prompt 由 `slice_interview_memory` 再筛。"""
     from app.core.dependencies import get_embedding_provider
-    from app.features.memory.evidence_format import format_evidence_chunks_for_prompt
     from app.features.memory.retrieval_trace import (
         chat_memory_retrieval_trace_from_bundle,
     )
@@ -69,13 +68,13 @@ async def _fetch_interview_memory_evidence(
         logger.debug(
             "event=chat_memory_retrieval_skip reason=disabled user_id={}", user_id
         )
-        return "", None
+        return None, None
     msg = (user_message or "").strip()
     if not msg:
         logger.debug(
             "event=chat_memory_retrieval_skip reason=empty user_id={}", user_id
         )
-        return "", None
+        return None, None
     try:
         emb = get_embedding_provider()
         ms = MemoryService(db, embedding_provider=emb)
@@ -85,30 +84,19 @@ async def _fetch_interview_memory_evidence(
         trace = chat_memory_retrieval_trace_from_bundle(
             bd, top_k=top_k, query_len=len(msg)
         )
-        text = format_evidence_chunks_for_prompt(bd)
-        t = (text or "").strip()
-        if not t:
-            logger.debug(
-                "event=memory_evidence_for_prompt user_id={} formatted_chars=0",
-                user_id,
-            )
-            return "", trace
-        max_c = settings.chat_memory_evidence_max_chars
-        if len(t) > max_c:
-            t = t[: max_c - 3] + "..."
         logger.info(
-            "event=memory_evidence_for_prompt user_id={} formatted_chars={}",
+            "event=memory_retrieval_bundle user_id={} top_k={}",
             user_id,
-            len(t),
+            top_k,
         )
-        return t, trace
+        return bd, trace
     except Exception as e:
         try:
             await db.rollback()
         except Exception as rollback_error:
             logger.warning("访谈记忆检索失败后回滚也失败: {}", rollback_error)
         logger.warning("访谈记忆检索失败: {}", e)
-        return "", None
+        return None, None
 
 
 class ChatOrchestrator:
@@ -284,14 +272,14 @@ class ChatOrchestrator:
             background_voice = infer_background_voice(user.occupation)
             occupation = user.occupation or ""
 
-        memory_evidence_text, mem_trace = await _fetch_interview_memory_evidence(
+        from app.features.memory.chat_memory_injection import slice_interview_memory
+
+        memory_bundle, mem_trace = await _fetch_interview_memory_bundle(
             db, user_id, normalized_user_message
         )
-        scene_cues = extract_scene_cues(normalized_user_message)
-        if scene_cues:
-            cue_block = "\n".join(f"- {c}" for c in scene_cues)
-            scene_hint = f"\n\n[场景氛围提示——可借用这些感官细节自然接话，不要原样抄]\n{cue_block}"
-            memory_evidence_text = (memory_evidence_text or "") + scene_hint
+        mem_slices = slice_interview_memory(memory_bundle, normalized_user_message)
+        # 场景关键词仅作为 focus planner 的辅助输入，不直接拼进记忆块，避免抢过用户明确的关系/身份线索
+        scene_cues_for_planner = extract_scene_cues(normalized_user_message)
 
         profile_birth_year = user.birth_year if user else None
         profile_era_place = ""
@@ -313,13 +301,16 @@ class ChatOrchestrator:
             memoir_state=prompt_state,
             user_profile_context=user_profile_context,
             detected_user_stage=detected,
-            memory_evidence_text=memory_evidence_text,
+            memory_evidence_text=mem_slices.prompt_excerpt,
+            memory_anchor_source=mem_slices.anchor_source,
+            memory_planner_text=mem_slices.planner_preview,
             background_voice=background_voice,
             normalized_user_message=normalized_user_message,
             occupation=occupation,
             profile_birth_year=profile_birth_year,
             profile_era_place=profile_era_place,
             stage_switched_this_turn=stage_switched_this_turn,
+            scene_cues_for_planner=scene_cues_for_planner,
         )
         recent_questions = prompt_state.recent_questions
         if turn.interview_state_meta and isinstance(turn.interview_state_meta, dict):
@@ -411,12 +402,15 @@ class ChatOrchestrator:
         audio_duration_seconds: int | None = None,
         detected_user_stage: str | None = None,
         memory_evidence_text: str = "",
+        memory_anchor_source: str = "",
+        memory_planner_text: str = "",
         background_voice: str = "default",
         normalized_user_message: str | None = None,
         occupation: str = "",
         profile_birth_year: int | None = None,
         profile_era_place: str = "",
         stage_switched_this_turn: bool = False,
+        scene_cues_for_planner: Optional[list[str]] = None,
     ) -> AgentChatTurn:
         """委托 InterviewAgent 生成访谈回复（持久化由调用方负责）。"""
         return await self.interview_agent.generate_response_with_state(
@@ -426,12 +420,15 @@ class ChatOrchestrator:
             user_profile_context=user_profile_context,
             detected_user_stage=detected_user_stage,
             memory_evidence_text=memory_evidence_text,
+            memory_anchor_source=memory_anchor_source,
+            memory_planner_text=memory_planner_text,
             background_voice=background_voice,
             normalized_user_message=normalized_user_message,
             occupation=occupation,
             profile_birth_year=profile_birth_year,
             profile_era_place=profile_era_place,
             stage_switched_this_turn=stage_switched_this_turn,
+            scene_cues_for_planner=scene_cues_for_planner,
         )
 
     def detect_user_stage(self, user_message: str) -> str: