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/tests/test_experience_regressions.py

@@ -4,6 +4,8 @@
 访谈侧仅验证 prompt 仍包含关键行为指引。
 """
 
+from app.agents.chat.interview_turn_plan import plan_interview_turn
+from app.agents.chat.output_rules import chat_output_rules
 from app.agents.chat.prompts_conversation import (
     get_guided_conversation_prompt,
     get_opening_prompt,
@@ -54,6 +56,23 @@ class TestChatExperienceRegressions:
         )
         assert "接住" in p
 
+    def test_guided_prompt_bans_assistant_autobiographical_claims(self) -> None:
+        """避免助手把用户经历说成自己的童年/暗恋等（身份越界）。"""
+        p = get_guided_conversation_prompt(
+            current_stage="education",
+            empty_slots=["school"],
+            filled_slots={},
+            detected_user_stage="education",
+            user_profile_context="",
+            persona="default",
+        )
+        assert "身份边界" in p
+        assert "暗恋" in p
+        assert "共同回忆" in p
+        rules = chat_output_rules()
+        assert "助手本人" in rules
+        assert "共同回忆" in rules
+
     def test_opening_prompt_stays_short_task_shape(self) -> None:
         p = get_opening_prompt(
             current_stage="childhood",
@@ -64,6 +83,19 @@ class TestChatExperienceRegressions:
         assert "问候" in p
         assert "任务" in p or "具体问题" in p
 
+    def test_followup_her_story_turn_plan_stays_user_perspective_shape(self) -> None:
+        """追问第三方故事时：编排仍锁用户视角，并倾向「承接+一问」形状（防助手自传漂移）。"""
+        p = plan_interview_turn(
+            current_stage="education",
+            empty_slots=["school", "city"],
+            normalized_user_message="讲讲她的故事吧，后来怎么样了？",
+            memory_evidence_text="",
+            stage_switched_this_turn=False,
+        )
+        assert p.subject_owner == "user_only"
+        assert p.forbid_first_person_experience is True
+        assert p.reply_shape == "ack_then_question"
+
 
 class TestMemoirStyleRegressions:
     """保护「回忆录有文笔」体验。"""

api/tests/test_interview_prompts.py

@@ -3,7 +3,9 @@
 from langchain_core.messages import AIMessage, HumanMessage, SystemMessage
 
 from app.agents.chat.interview_state_hints import (
+    AUTOBIOGRAPHICAL_BOUNDARY_FALLBACK_ZH,
     DUPLICATE_QUESTION_GUARD_FALLBACK_ZH,
+    apply_autobiographical_boundary_guard,
     apply_duplicate_question_guard,
     extract_scene_cues,
     segments_are_only_duplicate_guard_fallback,
@@ -16,6 +18,7 @@ from app.agents.state_schema import (
 )
 from app.agents.chat.helpers import format_history_string
 from app.agents.chat.personas import normalize_interview_persona
+from app.agents.chat.output_rules import chat_output_rules
 from app.agents.chat.prompts_conversation import (
     get_guided_conversation_prompt,
     get_opening_prompt,
@@ -217,8 +220,7 @@ def test_guided_prompt_contains_memory_section_when_evidence():
         persona="default",
         memory_evidence_text="[摘要:rolling] 1990年生于上海。",
     )
-    assert "相关记忆摘录" in p
-    assert "过往口述" in p
+    assert "记忆线索" in p or "追问角度" in p
     assert "1990年生于上海" in p
 
 
@@ -368,3 +370,92 @@ def test_format_history_string_omit_system_body() -> None:
     assert "System: <omitted total_len=16" in s
     assert "sha12=" in s
     assert "Human: hi" in s
+
+
+def test_guided_prompt_includes_identity_boundary_hard_rules() -> None:
+    p = get_guided_conversation_prompt(
+        current_stage="childhood",
+        empty_slots=["place"],
+        filled_slots={},
+        detected_user_stage="childhood",
+        user_profile_context="",
+        persona="default",
+    )
+    assert "身份边界" in p
+    assert "真实人生传记" in p
+    assert "共同回忆" in p
+    assert "泛指" in p
+
+
+def test_guided_prompt_blocks_using_user_context_as_assistant_identity() -> None:
+    p = get_guided_conversation_prompt(
+        current_stage="childhood",
+        empty_slots=["place"],
+        filled_slots={},
+        detected_user_stage="childhood",
+        user_profile_context="成长地：上海",
+        persona="default",
+    )
+    assert "我是上海人" in p
+    assert "不能把用户的成长地答成" in p
+    assert "你刚提到上海" in p or "你之前说过那段童年" in p
+
+
+def test_chat_output_rules_bans_assistant_autobiography() -> None:
+    rules = chat_output_rules()
+    assert "禁止" in rules
+    assert "声称助手本人" in rules or "助手本人" in rules
+    assert "共同回忆" in rules
+    assert "你是哪里人" in rules
+    assert "你刚提到" in rules
+
+
+def test_autobiographical_boundary_guard_replaces_crush_claim() -> None:
+    out, touched = apply_autobiographical_boundary_guard(
+        ["是啊，朱丽叶就是我当时暗恋的女生。"]
+    )
+    assert touched is True
+    assert out == [AUTOBIOGRAPHICAL_BOUNDARY_FALLBACK_ZH]
+
+
+def test_autobiographical_boundary_guard_replaces_childhood_claim() -> None:
+    out, touched = apply_autobiographical_boundary_guard(
+        ["我小时候也演过这个，还挺紧张的。"]
+    )
+    assert touched is True
+    assert out == [AUTOBIOGRAPHICAL_BOUNDARY_FALLBACK_ZH]
+
+
+def test_autobiographical_boundary_guard_allows_generic_empathy() -> None:
+    safe = [
+        "我能想象那会儿站在台上，手心里全是汗。",
+        "换作很多人可能也会记很久。",
+    ]
+    out, touched = apply_autobiographical_boundary_guard(safe)
+    assert touched is False
+    assert out == safe
+
+
+def test_autobiographical_boundary_guard_mixed_segments() -> None:
+    out, touched = apply_autobiographical_boundary_guard(
+        ["嗯，你刚才那段我接住了。", "我小时候也演过。"]
+    )
+    assert touched is True
+    assert out[0] == "嗯，你刚才那段我接住了。"
+    assert out[1] == AUTOBIOGRAPHICAL_BOUNDARY_FALLBACK_ZH
+
+
+def test_autobiographical_boundary_guard_catches_iyan_role_without_demo() -> None:
+    out, touched = apply_autobiographical_boundary_guard(
+        ["那次话剧里我演罗密欧，对手戏挺难的。"]
+    )
+    assert touched is True
+    assert out == [AUTOBIOGRAPHICAL_BOUNDARY_FALLBACK_ZH]
+
+
+def test_autobiographical_boundary_guard_allows_wo_yanshi_demo() -> None:
+    out, touched = apply_autobiographical_boundary_guard(
+        ["我演示一下这个按钮怎么点。"]
+    )
+    assert touched is False
+    assert out == ["我演示一下这个按钮怎么点。"]

api/tests/test_interview_turn_plan.py

@@ -1,7 +1,10 @@
 """interview_turn_plan：轮次模式与主槽选择（服务端硬编排）。"""
 
 from app.agents.chat.interview_turn_plan import (
+    InterviewTurnPlan,
+    apply_safe_mode_override,
     extract_anchor_snippet,
+    format_interview_turn_directive_block,
     plan_interview_turn,
     primary_empty_slot,
 )
@@ -12,11 +15,177 @@ def test_primary_empty_slot_order():
     assert primary_empty_slot("childhood", ["emotion"]) == "emotion"
 
 
-def test_extract_anchor_snippet_prefers_memory():
+def test_extract_anchor_snippet_prefers_user_when_long_enough():
     mem = "摘录的一段记忆\n\n[场景氛围提示"
-    assert "摘录的一段记忆" in extract_anchor_snippet(
-        memory_evidence_text=mem, user_message="用户说很长一句" * 3
+    um = "用户说很长一句" * 5
+    sn = extract_anchor_snippet(memory_evidence_text=mem, user_message=um)
+    assert sn.startswith("用户说")
+    assert "摘录" not in sn
+
+
+def test_extract_anchor_snippet_prefers_first_m_line():
+    mem = (
+        "【相关记忆摘录·聊天专用】\n"
+        "说明行……\n"
+        "[M1] 你在校园演出里饰演罗密欧。\n"
     )
+    sn = extract_anchor_snippet(memory_evidence_text=mem, user_message="短")
+    assert "校园演出" in sn
+    assert "【相关" not in sn
+
+
+def test_plan_sets_memory_usage_when_evidence_present():
+    p = plan_interview_turn(
+        current_stage="childhood",
+        empty_slots=["place"],
+        normalized_user_message="嗯。",
+        memory_evidence_text="【头】\n[M1] 你提过河边。",
+        stage_switched_this_turn=False,
+    )
+    assert p.memory_usage == "allowed_with_attribution"
+
+
+def test_plan_marks_assistant_identity_question() -> None:
+    p = plan_interview_turn(
+        current_stage="childhood",
+        empty_slots=["place"],
+        normalized_user_message="你是哪里人，你的童年是什么样的？",
+        memory_evidence_text="用户曾说：「我小时候在上海长大。」",
+        stage_switched_this_turn=False,
+    )
+    assert p.assistant_identity_question is True
+
+
+def test_plan_reply_shape_ack_then_question_on_her_story_followup():
+    p = plan_interview_turn(
+        current_stage="education",
+        empty_slots=["school"],
+        normalized_user_message="那你讲讲她的故事吧。",
+        memory_evidence_text="",
+        stage_switched_this_turn=False,
+    )
+    assert p.reply_shape == "ack_then_question"
+
+
+def test_directive_includes_attribution_when_memory_allowed():
+    plan = InterviewTurnPlan(
+        mode="memoir_push",
+        anchor_slot_key="place",
+        anchor_slot_readable="成长的地方",
+        anchor_snippet="挂钩",
+        memory_usage="allowed_with_attribution",
+        memory_reference_style="你之前提过",
+        reply_shape="ack_then_question",
+    )
+    block = format_interview_turn_directive_block(plan)
+    assert "线索" in block
+    assert "你之前提过" in block
+    assert "真实人生传记" in block
+    assert "ack" not in block.lower()
+    assert "承接" in block
+    assert (
+        "本轮追问" in block
+        or "承接角度" in block
+        or "本轮承接重点" in block
+    )
+
+
+def test_directive_includes_focus_summary_when_set():
+    plan = InterviewTurnPlan(
+        mode="memoir_push",
+        anchor_slot_key="place",
+        anchor_slot_readable="成长的地方",
+        anchor_snippet="挂钩",
+        focus_summary="先接住「芳芳」与怕丢脸这条线",
+        focus_source="llm",
+        primary_focus="relationship",
+    )
+    block = format_interview_turn_directive_block(plan)
+    assert "芳芳" in block
+    assert "本轮追问" in block or "承接角度" in block
+
+
+def test_directive_marks_user_message_anchor_source_correctly():
+    plan = InterviewTurnPlan(
+        mode="memoir_push",
+        anchor_slot_key="place",
+        anchor_slot_readable="成长的地方",
+        anchor_snippet="那年夏天我总往河边跑",
+        anchor_source_kind="user_message",
+    )
+    block = format_interview_turn_directive_block(plan)
+    assert "来自用户本轮原话摘录" in block
+    assert "不是用户本轮新说的内容" not in block
+
+
+def test_directive_marks_memory_anchor_source_correctly():
+    plan = InterviewTurnPlan(
+        mode="memoir_push",
+        anchor_slot_key="place",
+        anchor_slot_readable="成长的地方",
+        anchor_snippet="你小时候常在河边玩",
+        anchor_source_kind="memory",
+    )
+    block = format_interview_turn_directive_block(plan)
+    assert "来自检索到的用户过往口述/摘要" in block
+    assert "不是用户本轮新说的内容" in block
+
+
+def test_directive_adds_boundary_for_assistant_identity_question():
+    plan = InterviewTurnPlan(
+        mode="memoir_push",
+        anchor_slot_key="place",
+        anchor_slot_readable="成长的地方",
+        anchor_snippet="用户曾说：「我小时候在上海长大。」",
+        assistant_identity_question=True,
+    )
+    block = format_interview_turn_directive_block(plan)
+    assert "本轮用户在问助手本人" in block
+    assert "你刚提到上海" in block
+
+
+def test_apply_safe_mode_override_blocks_emotion_downgrade():
+    assert apply_safe_mode_override("emotion_first", "memoir_push", primary_focus="emotion") is None
+
+
+def test_apply_safe_mode_override_allows_memoir_to_emotion():
+    assert (
+        apply_safe_mode_override("memoir_push", "emotion_first", primary_focus="relationship")
+        == "emotion_first"
+    )
+
+
+def test_apply_safe_mode_override_clarify_blocks_memoir():
+    assert apply_safe_mode_override("clarify_first", "memoir_push", primary_focus="emotion") is None
+
+
+def test_apply_safe_mode_override_clarify_to_emotion():
+    assert (
+        apply_safe_mode_override("clarify_first", "emotion_first", primary_focus="emotion")
+        == "emotion_first"
+    )
+
+
+def test_plan_clarify_first_when_ambiguous():
+    p = plan_interview_turn(
+        current_stage="childhood",
+        empty_slots=["place", "people"],
+        normalized_user_message="那种喜欢我也说不清，好像有又好像没有。",
+        memory_evidence_text="",
+        stage_switched_this_turn=False,
+    )
+    assert p.mode == "clarify_first"
+
+
+def test_plan_clarify_first_when_very_short():
+    p = plan_interview_turn(
+        current_stage="childhood",
+        empty_slots=["place"],
+        normalized_user_message="还好吧",
+        memory_evidence_text="",
+        stage_switched_this_turn=False,
+    )
+    assert p.mode == "clarify_first"
 
 
 def test_plan_memoir_push():

api/tests/test_memory_evidence.py

@@ -3,6 +3,7 @@
 import pytest
 
 from app.features.memory import evidence as evidence_mod
+from app.features.memory.evidence_format import format_evidence_chunks_for_chat_prompt
 from app.features.memory.evidence import (
     EMPTY_EVIDENCE_BUNDLE,
     _facts_to_dicts,
@@ -85,3 +86,107 @@ def test_format_helpers_empty() -> None:
     assert _facts_to_dicts([]) == []
     assert _timeline_to_dicts([]) == []
     assert _stories_to_dicts([]) == []
+
+
+def test_format_evidence_chunks_for_chat_prompt_reframes_and_labels() -> None:
+    evidence = {
+        "relevant_chunks": [
+            {"id": "chunk-1", "content": "我小时候在河边长大，夏天常去玩水。"},
+        ],
+        "relevant_summaries": [],
+        "relevant_facts": [],
+        "timeline_hints": [],
+        "relevant_stories": [],
+    }
+    text = format_evidence_chunks_for_chat_prompt(evidence)
+    assert "聊天专用" in text
+    assert "归因" in text
+    assert "[M1]" in text
+    assert "用户曾说" in text
+    assert "我小时候在河边长大" in text
+
+
+def test_slice_interview_memory_empty_bundle():
+    from app.features.memory.chat_memory_injection import slice_interview_memory
+
+    s = slice_interview_memory(None, "你好")
+    assert s.prompt_excerpt == ""
+    assert s.anchor_source == ""
+    assert s.planner_preview == ""
+    assert s.had_retrieval is False
+
+
+def test_slice_interview_memory_retrieval_not_equal_inject_dismissive():
+    """有检索预览但 gating 后不进主 prompt / anchor。"""
+    from app.features.memory.chat_memory_injection import slice_interview_memory
+
+    evidence = {
+        "relevant_chunks": [
+            {"id": "c1", "content": "很久以前在校园礼堂排练到很晚。"},
+        ],
+        "relevant_summaries": [],
+        "relevant_facts": [],
+        "timeline_hints": [],
+        "relevant_stories": [],
+    }
+    s = slice_interview_memory(evidence, "哈哈，早就不会了")
+    assert s.prompt_excerpt == ""
+    assert s.anchor_source == ""
+    assert s.planner_preview.strip() != ""
+    assert s.had_retrieval is True
+
+
+def test_slice_interview_memory_minimal_inject_when_aligned():
+    from app.features.memory.chat_memory_injection import slice_interview_memory
+
+    evidence = {
+        "relevant_chunks": [
+            {"id": "c1", "content": "你在校园演出里饰演罗密欧。"},
+        ],
+        "relevant_summaries": [],
+        "relevant_facts": [],
+        "timeline_hints": [],
+        "relevant_stories": [],
+    }
+    s = slice_interview_memory(evidence, "那次排练其实挺紧张的，灯光一打我就忘词。")
+    assert "记忆线索" in s.prompt_excerpt
+    assert "校园演出" in s.prompt_excerpt or "罗密欧" in s.prompt_excerpt
+    assert s.anchor_source
+    assert s.had_retrieval is True
+
+
+def test_slice_interview_memory_keeps_first_person_but_marks_ownership():
+    from app.features.memory.chat_memory_injection import slice_interview_memory
+
+    evidence = {
+        "relevant_chunks": [
+            {"id": "c1", "content": "我小时候在河边长大，夏天常去玩水。"},
+        ],
+        "relevant_summaries": [],
+        "relevant_facts": [],
+        "timeline_hints": [],
+        "relevant_stories": [],
+    }
+    s = slice_interview_memory(evidence, "那条河一到夏天就特别热闹，我现在都记得。")
+    assert "用户曾说" in s.prompt_excerpt
+    assert "我小时候在河边长大" in s.prompt_excerpt
+    assert s.anchor_source.startswith("用户曾说")
+
+
+def test_slice_interview_memory_suppresses_long_new_topic():
+    from app.features.memory.chat_memory_injection import slice_interview_memory
+
+    evidence = {
+        "relevant_chunks": [
+            {"id": "c1", "content": "旧记忆关于河边。"},
+        ],
+        "relevant_summaries": [],
+        "relevant_facts": [],
+        "timeline_hints": [],
+        "relevant_stories": [],
+    }
+    long_msg = "我今天想随便聊聊工作里的事，项目压力很大。" * 6
+    assert len(long_msg) > 72
+    s = slice_interview_memory(evidence, long_msg)
+    assert s.prompt_excerpt == ""
+    assert s.anchor_source == ""

api/tests/test_reply_planner.py

@@ -0,0 +1,112 @@
+"""reply_planner：JSON 合并与安全边界。"""
+
+import json
+
+from app.agents.chat.interview_turn_plan import InterviewTurnPlan
+from app.agents.chat.reply_planner import merge_reply_planner_json_into_turn_plan
+
+
+def _base_plan(**kwargs) -> InterviewTurnPlan:
+    defaults = dict(
+        mode="memoir_push",
+        anchor_slot_key="place",
+        anchor_slot_readable="成长的地方",
+        anchor_snippet="河边",
+        memory_usage="allowed_with_attribution",
+        reply_shape="flexible",
+        memory_reference_style="你之前提过",
+    )
+    defaults.update(kwargs)
+    return InterviewTurnPlan(**defaults)
+
+
+def test_merge_does_not_upgrade_memory_from_none_to_allowed():
+    plan = _base_plan(memory_usage="none")
+    raw = json.dumps(
+        {
+            "memory_usage": "allowed_with_attribution",
+            "reply_shape": "ack_then_question",
+            "memory_reference_style": "你说过",
+            "forbid_first_person_experience": True,
+        }
+    )
+    merged = merge_reply_planner_json_into_turn_plan(plan, raw)
+    assert merged.memory_usage == "none"
+    assert merged.reply_shape == "ack_then_question"
+    assert merged.memory_reference_style == "你说过"
+
+
+def test_merge_allows_downgrade_memory_usage_to_none():
+    plan = _base_plan(memory_usage="allowed_with_attribution")
+    raw = json.dumps({"memory_usage": "none"})
+    merged = merge_reply_planner_json_into_turn_plan(plan, raw)
+    assert merged.memory_usage == "none"
+
+
+def test_merge_invalid_json_returns_original():
+    plan = _base_plan()
+    merged = merge_reply_planner_json_into_turn_plan(plan, "not json")
+    assert merged == plan
+
+
+def test_merge_ignores_non_dict_json():
+    plan = _base_plan()
+    merged = merge_reply_planner_json_into_turn_plan(plan, "[1,2]")
+    assert merged == plan
+
+
+def test_merge_trims_memory_reference_style():
+    plan = _base_plan()
+    raw = json.dumps({"memory_reference_style": "  你刚讲到  "})
+    merged = merge_reply_planner_json_into_turn_plan(plan, raw)
+    assert merged.memory_reference_style == "你刚讲到"
+
+
+def test_merge_sets_focus_and_summary():
+    plan = _base_plan()
+    raw = json.dumps(
+        {
+            "primary_focus": "relationship",
+            "focus_summary": "先接住用户提到的在场关系与面子压力",
+            "forbid_first_person_experience": True,
+        }
+    )
+    merged = merge_reply_planner_json_into_turn_plan(plan, raw)
+    assert merged.primary_focus == "relationship"
+    assert "关系" in merged.focus_summary
+    assert merged.focus_source == "llm"
+
+
+def test_merge_mode_override_memoir_to_emotion_when_focus_supports():
+    plan = _base_plan(mode="memoir_push")
+    raw = json.dumps(
+        {
+            "mode_override": "emotion_first",
+            "primary_focus": "identity",
+            "forbid_first_person_experience": True,
+        }
+    )
+    merged = merge_reply_planner_json_into_turn_plan(plan, raw)
+    assert merged.mode == "emotion_first"
+    assert merged.focus_source == "llm"
+
+
+def test_merge_rejects_unsafe_mode_override_emotion_to_memoir():
+    plan = _base_plan(mode="emotion_first")
+    raw = json.dumps(
+        {
+            "mode_override": "memoir_push",
+            "primary_focus": "memoir_gap",
+            "forbid_first_person_experience": True,
+        }
+    )
+    merged = merge_reply_planner_json_into_turn_plan(plan, raw)
+    assert merged.mode == "emotion_first"
+
+
+def test_merge_omitted_secondary_focus_unchanged():
+    plan = _base_plan()
+    raw = json.dumps({"reply_shape": "ack_only", "forbid_first_person_experience": True})
+    merged = merge_reply_planner_json_into_turn_plan(plan, raw)
+    assert merged.reply_shape == "ack_only"
+    assert merged.focus_source == "rule"