fix(sdk): placeholder values never overwrite the basic profile

A weak extraction model filled basic_info with "用户", "The user" or "N/A"
when the user stated nothing, and each placeholder superseded the real L0
value. upsert_basic_profile now drops placeholder values and writes no new
version when nothing is left; profile assembly skips stored placeholders.
The writer keeps a basic_info field only when its value appears in a user
message of the request. The zh and en extraction prompts add a hard rule
and a small-talk counter-example.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

hy_memory/agent/extractor.py

@@ -682,6 +682,23 @@ Output:
 ```
 John's occupation and location are HIS attributes, so they go into `memory` with "John" as the subject. They must NEVER land in `basic_info.occupation` / `basic_info.location` — `basic_info` is exclusively the user's own profile. Only `name` (Tim, from the user message's `name` field) belongs there.
 
+## Example 11 — Counter-example: small talk has no basic_info
+
+New Messages:
+[{"role": "user", "content": "Hi"},
+ {"role": "assistant", "content": "Hello! The user's name is The user. How can I help?"}]
+Memory Date: 2025-09-03
+
+Output:
+```json
+{
+  "memory": [],
+  "intentions": [],
+  "basic_info": {}
+}
+```
+The user stated nothing about themselves. The assistant's line is not a user turn, and "The user" is a placeholder, not a name. Output `{}` — never `{"name": "The user"}` or `{"name": "user"}`.
+
 ---
 """
 
@@ -749,6 +766,14 @@ Rules:
   fields (e.g. name, age, location, occupation, employer) **explicitly stated**
   in the NEW user messages. Never infer, never invent. If no field is explicitly
   stated, output `{{}}`.
+  * HARD RULE: fill a field ONLY if the user stated it in a `user` turn.
+    Anything said only by the assistant (including the assistant quoting an
+    earlier memory such as "The user's name is X") does NOT count. If nothing
+    qualifies, output `"basic_info": {{}}`.
+  * NEVER write the field name itself or a placeholder as a value ("user",
+    "The user", "用户", "name", "N/A", "unknown", "none", "-"). Omit the field.
+  * Negative example: small talk such as `[user]: Hi` / `[assistant]: How can
+    I help?`, or `[user]: Reply with exactly: PONG` → `"basic_info": {{}}`.
   * Exception for `name`: the `name` field on the user's own message, or a leading
     line "The user's name is X" / "用户名叫X", is an authoritative source for
     `name` — even though the user never said it out loud.
@@ -817,6 +842,14 @@ Rules:
   fields (e.g. name, age, location, occupation, employer) **explicitly stated**
   in the NEW user messages. Never infer, never invent. If no field is explicitly
   stated, output `{{}}`.
+  * HARD RULE: fill a field ONLY if the user stated it in a `user` turn.
+    Anything said only by the assistant (including the assistant quoting an
+    earlier memory such as "The user's name is X") does NOT count. If nothing
+    qualifies, output `"basic_info": {{}}`.
+  * NEVER write the field name itself or a placeholder as a value ("user",
+    "The user", "用户", "name", "N/A", "unknown", "none", "-"). Omit the field.
+  * Negative example: small talk such as `[user]: Hi` / `[assistant]: How can
+    I help?`, or `[user]: Reply with exactly: PONG` → `"basic_info": {{}}`.
   * Exception for `name`: the `name` field on the user's own message, or a leading
     line "The user's name is X" / "用户名叫X", is an authoritative source for
     `name` — even though the user never said it out loud.

hy_memory/agent/prompts_zh.py

@@ -365,6 +365,13 @@ EXTRACT_PROMPT_ZH = """## 最近 k 条消息
 * basic_info：**只装用户本人的画像**。
   仅当 NEW user messages 中**明确**陈述稳定画像字段（如 姓名、年龄、所在地、职业、雇主）时才填写。
   绝不推断、绝不编造。没有任何字段被明确陈述时输出 `{{}}`。
+  * **硬性规则**：只有用户在 `user` 回合里**明确说出**的字段才能填写。只在
+    assistant 回合出现的内容（包括 assistant 引用旧记忆，如"The user's name is X"
+    / "用户名是X"）一律不算。没有符合条件的字段时，必须输出 `"basic_info": {{}}`。
+  * **绝不**把字段名本身或占位符当值写入（"用户"、"The user"、"user"、"name"、
+    "姓名"、"N/A"、"unknown"、"未知"、"未提供"、"-"），宁可省略该字段。
+  * 负例：寒暄如 `[user]: 你好` / `[assistant]: 我可以帮你。`，或
+    `[user]: Reply with exactly: PONG` → `"basic_info": {{}}`。
   * `name` 的例外：用户自己那条消息上的 `name` 字段，或文首形如
     "用户名叫X" / "The user's name is X" 的那一行，都是 `name` 的权威来源
     ——即使用户本人从未开口自报姓名。
@@ -440,6 +447,13 @@ EXTRACT_PROMPT_ZH_NO_INTENTION = """## 最近 k 条消息
 * basic_info：**只装用户本人的画像**。
   仅当 NEW user messages 中**明确**陈述稳定画像字段（如 姓名、年龄、所在地、职业、雇主）时才填写。
   绝不推断、绝不编造。没有任何字段被明确陈述时输出 `{{}}`。
+  * **硬性规则**：只有用户在 `user` 回合里**明确说出**的字段才能填写。只在
+    assistant 回合出现的内容（包括 assistant 引用旧记忆，如"The user's name is X"
+    / "用户名是X"）一律不算。没有符合条件的字段时，必须输出 `"basic_info": {{}}`。
+  * **绝不**把字段名本身或占位符当值写入（"用户"、"The user"、"user"、"name"、
+    "姓名"、"N/A"、"unknown"、"未知"、"未提供"、"-"），宁可省略该字段。
+  * 负例：寒暄如 `[user]: 你好` / `[assistant]: 我可以帮你。`，或
+    `[user]: Reply with exactly: PONG` → `"basic_info": {{}}`。
   * `name` 的例外：用户自己那条消息上的 `name` 字段，或文首形如
     "用户名叫X" / "The user's name is X" 的那一行，都是 `name` 的权威来源
     ——即使用户本人从未开口自报姓名。
@@ -669,6 +683,23 @@ Memory Date：2025-09-02
 ```
 职业和所在地是 **John 的**属性，所以进 `memory` 并以"John"作主体，**绝不能**落到 `basic_info` 的 `occupation` / `location` —— `basic_info` 专属于用户本人。只有 `name`（Tim，来自用户消息的 `name` 字段）才该放进去。注意输出语言随输入语言（此例为英文）。
 
+## 示例 11 —— 反例：寒暄没有 basic_info
+
+新消息：
+[{"role": "user", "content": "你好"},
+ {"role": "assistant", "content": "你好！我可以帮你。之前记得 The user's name is 用户."}]
+Memory Date：2025-09-03
+
+输出：
+```json
+{
+  "memory": [],
+  "intentions": [],
+  "basic_info": {}
+}
+```
+用户没有说任何关于自己的信息。assistant 那句不是 user 回合，"用户" 也是占位符而不是名字。必须输出 `{}`——**绝不能**输出 `{"name": "用户"}`。
+
 ---
 """
 

hy_memory/agent/tools/basic_profile.py

@@ -26,6 +26,8 @@ from __future__ import annotations
 
 import asyncio
 import logging
+import re
+import unicodedata
 import uuid
 import weakref
 from dataclasses import dataclass
@@ -208,6 +210,147 @@ def _sanitize_kv(
     return result
 
 
+# ================================================================
+# 占位符守卫（P0-1）
+# ================================================================
+# 弱模型（如 qwen3.5-2b）在用户什么都没说时，会把 schema 字段说明里的词当值
+# 填进 basic_info：{"name": "用户"} / "The user" / "N/A"。每个占位符都会
+# supersede 掉真实值，演化链变成 费天翔 → Fei → 用户 → The user。
+# 规则：占位符永远不入库；读侧组装 KV 时占位符也不覆盖已有真实值。
+
+_PLACEHOLDER_VALUES = frozenset({
+    "用户", "该用户", "本用户", "用户本人",
+    "the user", "the user's", "user", "a user", "this user",
+    "n/a", "na", "none", "null", "nil", "unknown", "unspecified",
+    "not provided", "not specified", "not mentioned",
+    "未提供", "未知", "不详", "未说明", "未提及", "无",
+    "-", "?",
+})
+
+# 内置字段的中文标签：弱模型也会把字段说明的中文词当值。
+_FIELD_LABELS: Dict[str, tuple] = {
+    "name": ("姓名", "名字", "用户名", "用户姓名", "用户名字"),
+    "age": ("年龄", "用户年龄"),
+    "location": ("所在地", "地点", "城市", "用户所在地"),
+    "occupation": ("职业", "工作", "用户职业"),
+    "employer": ("雇主", "公司", "用户雇主"),
+}
+
+# 去掉标点/符号后再比较，让 "The user." / "用户。" / "N/A" 都命中。
+# "-" / "?" 本身就是纯标点，剥完为空串，同样判为占位符。
+_PUNCT_RE = re.compile(r"[\s\.,;:!?'\"`~()\[\]{}<>/\\|_\-\u3000-\u303f\uff00-\uff0f\uff1a-\uff20]+")
+
+
+def _placeholder_key(value: str) -> str:
+    s = unicodedata.normalize("NFKC", str(value or "")).strip().lower()
+    s = s.replace("\u2019", "'")
+    s = _PUNCT_RE.sub(" ", s)
+    return " ".join(s.split())
+
+
+_PLACEHOLDER_KEYS = frozenset(_placeholder_key(v) for v in _PLACEHOLDER_VALUES)
+
+
+def is_placeholder_value(field_name: str, value: Any) -> bool:
+    """value 是否为占位符（空 / 用户 / The user / N/A / 字段名本身 等）。
+
+    大小写不敏感、首尾空白与标点不敏感。数字（如 age=30）永远不是占位符。
+    """
+    if value is None:
+        return True
+    if isinstance(value, bool):
+        return False
+    if isinstance(value, (int, float)):
+        return False
+    key = _placeholder_key(value)
+    if not key:
+        return True
+    if key in _PLACEHOLDER_KEYS:
+        return True
+    field_key = _placeholder_key(field_name)
+    if field_key and key in (field_key, field_key.replace("_", " ")):
+        return True
+    for label in _FIELD_LABELS.get(field_name, ()):
+        if key == _placeholder_key(label):
+            return True
+    return False
+
+
+def _drop_placeholders(kv: Dict[str, Any]) -> Dict[str, Any]:
+    return {k: v for k, v in kv.items() if not is_placeholder_value(k, v)}
+
+
+# ================================================================
+# 来源守卫（P0-1）：basic_info 只认 user 回合
+# ================================================================
+# 抽取器偶尔把 assistant 回合里的内容（包括 assistant 引用的旧占位记忆
+# "The user's name is 用户."）当成用户画像再写一次。writer 把 user 回合原文
+# 交给 filter_basic_info_by_user_text()，值不在 user 原文里的字段一律丢弃。
+#
+# 匹配规则（大小写、空白不敏感，NFKC 归一）：
+# - 严格字段（name / age / location / employer）：值必须作为子串出现在 user
+#   原文里。"我叫费天翔" → name=费天翔 通过；"Fei Tianxiang" 与 "fei  tianxiang"
+#   通过；"杭州" 写成 "Hangzhou" 则丢弃（宁缺毋滥：身份字段错一次就会
+#   supersede 掉真实值）。
+# - 其余字段（occupation 及业务自定义字段）：允许模型轻度改写。只要值与 user
+#   原文的最长公共子串 ≥ 2 个字符且覆盖值长度的一半以上即通过，例如
+#   "AI 基础设施从业者" vs "做 AI 基础设施"（公共子串 "ai基础设施" 覆盖 6/9）。
+#   完全凭空的值（与 user 原文无公共片段）仍被丢弃。
+
+STRICT_SOURCE_FIELDS = frozenset({"name", "age", "location", "employer"})
+
+
+def _match_key(text: Any) -> str:
+    s = unicodedata.normalize("NFKC", str(text if text is not None else "")).lower()
+    s = s.replace("\u2019", "'")
+    return "".join(s.split())
+
+
+def _longest_common_substring(a: str, b: str) -> int:
+    if not a or not b:
+        return 0
+    best = 0
+    prev = [0] * (len(b) + 1)
+    for ca in a:
+        cur = [0] * (len(b) + 1)
+        for j, cb in enumerate(b, 1):
+            if ca == cb:
+                cur[j] = prev[j - 1] + 1
+                if cur[j] > best:
+                    best = cur[j]
+        prev = cur
+    return best
+
+
+def value_in_user_text(field_name: str, value: Any, user_text: str) -> bool:
+    """value 是否出自 user 原文（规则见上方注释）。"""
+    v = _match_key(value)
+    t = _match_key(user_text)
+    if not v or not t:
+        return False
+    if v in t:
+        return True
+    if field_name in STRICT_SOURCE_FIELDS:
+        return False
+    common = _longest_common_substring(v, t)
+    return common >= 2 and common * 2 >= len(v)
+
+
+def filter_basic_info_by_user_text(
+    kv: Dict[str, Any],
+    user_text: Optional[str],
+) -> Dict[str, Any]:
+    """只保留值出现在 user 原文里的字段；user_text 为空时整份丢弃。"""
+    if not isinstance(kv, dict) or not kv:
+        return {}
+    if not (user_text or "").strip():
+        return {}
+    return {
+        k: v for k, v in kv.items()
+        if value_in_user_text(k, v, user_text or "")
+    }
+
+
 # ================================================================
 # 公共入口：upsert_basic_profile
 # ================================================================
@@ -287,6 +430,8 @@ async def _upsert_basic_profile_unlocked(
     # L0 Basic Info 从输入、唯一性到落盘都只属于 user_id。
 
     new_kv = _sanitize_kv(kv or {}, allowed_fields=allowed_fields)
+    # 单调性守卫：占位符永远不覆盖已有值，过滤后为空就不写新版本、不产生新链头。
+    new_kv = _drop_placeholders(new_kv)
     if not new_kv:
         return BasicProfileUpsertResult(
             success=True, changed=False, reason="no valid fields after sanitize"
@@ -513,5 +658,8 @@ def _assemble_full_kv_from_nodes(
         for k, v in kv.items():
             if allowed_set is not None and k not in allowed_set:
                 continue
+            # 历史上已落盘的占位符版本（如 name=用户）不覆盖更早的真实值。
+            if is_placeholder_value(k, v):
+                continue
             full_kv[k] = v
     return full_kv

hy_memory/pipelines/writer.py

@@ -41,7 +41,10 @@ from ..core.workload import (
 )
 from ..agent.mem_agent import MemAgent, ProcessMode
 from ..agent.reconciler import MemoryReconciler
-from ..agent.tools.basic_profile import upsert_basic_profile
+from ..agent.tools.basic_profile import (
+    filter_basic_info_by_user_text,
+    upsert_basic_profile,
+)
 from ..models.memory import MemoryNode, MemoryLayer, MemoryStatus, SourceType
 from ..data.vector_store import create_vector_store
 from ..data.vector_store_base import VectorStoreBase
@@ -3305,6 +3308,20 @@ class MemoryWriter(WritePipeline):
         tool_calls_raw = None  # v0.1.5.13+ 不再有真实 LLM tool_calls；保留字段做兼容
         tool_results_summary: List[Dict[str, Any]] = []
 
+        # 来源守卫：basic_info 只认 user 回合。值不在任何 user 消息里的字段
+        # （例如 assistant 引用旧占位记忆 "The user's name is 用户."）一律丢弃；
+        # 有 messages 但没有 user 消息时整份丢弃。纯文本 add（无 messages）的
+        # content 即调用方给出的用户原文，按 user 回合处理。
+        if isinstance(basic_info_raw, dict) and basic_info_raw:
+            _user_text = self._basic_info_user_text(request)
+            _kept = filter_basic_info_by_user_text(basic_info_raw, _user_text)
+            if len(_kept) != len(basic_info_raw):
+                logger.info(
+                    f"[write] basic_info dropped fields not stated in user turns: "
+                    f"{sorted(set(basic_info_raw) - set(_kept))}"
+                )
+            basic_info_raw = _kept
+
         if isinstance(basic_info_raw, dict) and basic_info_raw:
             try:
                 _bp_result = await upsert_basic_profile(
@@ -3580,6 +3597,27 @@ class MemoryWriter(WritePipeline):
             "llm_total_tokens": getattr(_er, "tokens_used", 0) or 0,
         }
 
+    @staticmethod
+    def _basic_info_user_text(request) -> str:
+        """basic_info 来源校验用的 user 原文。
+
+        - 有 messages：只取 role=user 的 content 与 user 自带的 ``name``（name 是
+          basic_info.name 的权威来源）；没有 user 消息时返回空串（整份丢弃）。
+        - 无 messages（纯文本 add）：content 本身就是用户原文。
+        """
+        messages = getattr(request, "messages", None) or []
+        if not messages:
+            return str(getattr(request, "content", "") or "")
+        parts: List[str] = []
+        for m in messages:
+            if getattr(m, "role", None) != "user":
+                continue
+            name = getattr(m, "name", None)
+            if name:
+                parts.append(str(name))
+            parts.append(str(getattr(m, "content", "") or ""))
+        return "\n".join(parts)
+
     @staticmethod
     def _mark_agent_failure(
         response: WriteResponse,

tests/test_basic_info_placeholder_guard.py

@@ -0,0 +1,238 @@
+"""P0-1：占位符不得污染 L0 基础画像。
+
+弱模型（qwen3.5-2b）在用户什么都没说时把 basic_info 填成 {"name": "用户"} /
+"The user" / "N/A"，每个占位符都 supersede 掉真实值，演化链变成
+费天翔 → Fei → 用户 → The user。三层防御：
+(a) upsert 单调性守卫；(b) writer 只认 user 回合；(c) prompt 硬约束。
+本文件锁住 (a)(b) 与 (c) 的 prompt 文案，覆盖改进建议第 10 节验收 1–4。
+"""
+import asyncio
+from datetime import datetime, timedelta
+
+import pytest
+
+from hy_memory.agent.extractor import EXTRACT_FEW_SHOT_EN
+from hy_memory.agent.prompts_zh import EXTRACT_FEW_SHOT_ZH
+from hy_memory.agent.tools.basic_profile import (
+    _assemble_full_kv_from_nodes,
+    filter_basic_info_by_user_text,
+    is_placeholder_value,
+    upsert_basic_profile,
+    value_in_user_text,
+)
+from hy_memory.models.memory import MemoryLayer, MemoryNode, MemoryStatus
+from hy_memory.pipelines.base import ChatMessage, WriteRequest
+from hy_memory.pipelines.writer import MemoryWriter
+
+FIELDS = ["name", "age", "location", "occupation", "employer"]
+
+
+class FakeStore:
+    def __init__(self, nodes=None):
+        self.nodes = list(nodes or [])
+
+    async def list_by_user(self, **kwargs):
+        return [n for n in self.nodes if n.user_id == kwargs.get("user_id")]
+
+    async def get_by_id(self, node_id):
+        return next((n for n in self.nodes if n.node_id == node_id), None)
+
+    async def upsert(self, node):
+        self.nodes.append(node)
+
+    async def update_payload(self, node_id, updates):
+        node = await self.get_by_id(node_id)
+        for key, value in updates.items():
+            if key == "status" and isinstance(value, str):
+                value = MemoryStatus(value)
+            setattr(node, key, value)
+
+
+class FakeEmbed:
+    async def embed_queued(self, text):
+        return [0.1] * 8
+
+
+def l0_versions(store):
+    return [n for n in store.nodes if n.layer == MemoryLayer.L0_BASIC_INFO]
+
+
+def write_basic_info(store, request, llm_basic_info):
+    """模拟 writer 的 basic_info 段：来源守卫 → upsert。"""
+    user_text = MemoryWriter._basic_info_user_text(request)
+    kv = filter_basic_info_by_user_text(llm_basic_info, user_text)
+    if not kv:
+        return None
+    return asyncio.run(upsert_basic_profile(
+        user_id=request.user_id, kv=kv, vector_store=store,
+        embed_service=FakeEmbed(), allowed_fields=FIELDS,
+    ))
+
+
+def req(*msgs, content=""):
+    return WriteRequest(
+        content=content,
+        messages=[ChatMessage(role=r, content=c) for r, c in msgs],
+        user_id="u1",
+    )
+
+
+# ── 验收 1：寒暄不产生新的 L0 版本 ──
+
+def test_small_talk_creates_no_l0_version():
+    store = FakeStore()
+    r = req(("user", "你好"), ("assistant", "我可以帮你。"))
+    write_basic_info(store, r, {"name": "用户"})
+    write_basic_info(store, r, {})
+    assert l0_versions(store) == []
+
+
+def test_placeholder_only_upsert_writes_nothing():
+    store = FakeStore()
+    for kv in ({"name": "用户"}, {"name": "The user"}, {"name": "N/A"},
+               {"location": "unknown", "occupation": "未提供"}, {"name": "name"}):
+        res = asyncio.run(upsert_basic_profile(
+            user_id="u1", kv=kv, vector_store=store,
+            embed_service=FakeEmbed(), allowed_fields=FIELDS,
+        ))
+        assert res.success and not res.changed
+    assert l0_versions(store) == []
+
+
+# ── 验收 2：assistant 引用 "name is 用户" 不写入 ──
+
+def test_assistant_quoting_placeholder_does_not_write():
+    store = FakeStore()
+    r = req(
+        ("user", "你看那条记忆"),
+        ("assistant", "那条 The user's name is 用户. 更说明问题"),
+    )
+    write_basic_info(store, r, {"name": "用户"})
+    assert l0_versions(store) == []
+
+
+def test_value_only_in_assistant_turn_is_dropped():
+    r = req(("user", "帮我查下天气"), ("assistant", "好的，费天翔，杭州今天晴。"))
+    text = MemoryWriter._basic_info_user_text(r)
+    assert filter_basic_info_by_user_text(
+        {"name": "费天翔", "location": "杭州"}, text
+    ) == {}
+
+
+def test_no_user_messages_drops_basic_info():
+    r = req(("assistant", "我叫费天翔，常住杭州"))
+    assert MemoryWriter._basic_info_user_text(r) == ""
+    assert filter_basic_info_by_user_text({"name": "费天翔"}, "") == {}
+
+
+# ── 验收 3：用户自报姓名 ──
+
+def test_user_states_name():
+    store = FakeStore()
+    r = req(("user", "我叫费天翔，常住杭州，做 AI 基础设施。"))
+    res = write_basic_info(store, r, {
+        "name": "费天翔", "location": "杭州", "occupation": "AI 基础设施从业者",
+    })
+    assert res.success and res.created
+    assert l0_versions(store)[0].custom["basic_info_kv"] == {
+        "name": "费天翔", "location": "杭州", "occupation": "AI 基础设施从业者",
+    }
+
+
+def test_user_message_name_attribute_counts_as_user_text():
+    r = WriteRequest(
+        messages=[ChatMessage(role="user", name="Tim", content="What's up?")],
+        user_id="u1",
+    )
+    text = MemoryWriter._basic_info_user_text(r)
+    assert filter_basic_info_by_user_text({"name": "Tim"}, text) == {"name": "Tim"}
+
+
+def test_plain_text_add_is_user_text():
+    r = WriteRequest(content="I'm Fei Tianxiang, 35, living in Hangzhou.", user_id="u1")
+    text = MemoryWriter._basic_info_user_text(r)
+    assert filter_basic_info_by_user_text(
+        {"name": "fei  tianxiang", "age": 35, "location": "Hangzhou"}, text
+    ) == {"name": "fei  tianxiang", "age": 35, "location": "Hangzhou"}
+
+
+def test_strict_fields_require_containment_loose_fields_allow_rewording():
+    text = "我叫费天翔，常住杭州，做 AI 基础设施。"
+    assert value_in_user_text("name", "费天翔", text)
+    assert not value_in_user_text("location", "Hangzhou", text)
+    assert value_in_user_text("occupation", "AI 基础设施从业者", text)
+    assert not value_in_user_text("occupation", "软件工程师", text)
+
+
+# ── 验收 4：已有真实值不被后续占位符覆盖 ──
+
+def test_existing_name_not_overwritten_by_placeholder():
+    store = FakeStore()
+    write_basic_info(store, req(("user", "我叫费天翔")), {"name": "费天翔"})
+    assert len(l0_versions(store)) == 1
+    for kv in ({"name": "用户"}, {"name": "The user."}, {"name": "N/A"}):
+        res = asyncio.run(upsert_basic_profile(
+            user_id="u1", kv=kv, vector_store=store,
+            embed_service=FakeEmbed(), allowed_fields=FIELDS,
+        ))
+        assert not res.changed
+    assert len(l0_versions(store)) == 1
+    assert _assemble_full_kv_from_nodes(list(store.nodes)) == {"name": "费天翔"}
+
+
+def test_mixed_kv_keeps_real_fields_and_drops_placeholders():
+    store = FakeStore()
+    res = asyncio.run(upsert_basic_profile(
+        user_id="u1", kv={"name": "用户", "location": "杭州"}, vector_store=store,
+        embed_service=FakeEmbed(), allowed_fields=FIELDS,
+    ))
+    assert res.diff_kv == {"location": "杭州"}
+
+
+def test_already_poisoned_chain_heals_on_read():
+    now = datetime.utcnow()
+
+    def node(nid, kv, days_ago):
+        return MemoryNode(
+            node_id=nid, user_id="u1", agent_id="",
+            layer=MemoryLayer.L0_BASIC_INFO, content="profile",
+            custom={"basic_info_kv": kv},
+            gmt_created=now - timedelta(days=days_ago),
+        )
+
+    chain = [
+        node("a", {"name": "费天翔"}, 4), node("b", {"name": "Fei"}, 3),
+        node("c", {"name": "用户"}, 2), node("d", {"name": "The user"}, 1),
+    ]
+    assert _assemble_full_kv_from_nodes(chain) == {"name": "Fei"}
+
+
+@pytest.mark.parametrize("value", [
+    "", "  ", "用户", "用户。", "The user", "the user.", "THE USER", "user",
+    "The user's", "n/a", "N/A", "na", "None", "null", "unknown", "未提供",
+    "未知", "不详", "-", "?", "name", " Name ", "姓名",
+])
+def test_placeholder_values(value):
+    assert is_placeholder_value("name", value)
+
+
+@pytest.mark.parametrize("field,value", [
+    ("name", "费天翔"), ("name", "Fei"), ("name", "Username42"),
+    ("age", 30), ("age", 0), ("location", "杭州"), ("employer", "Tencent"),
+])
+def test_real_values_are_not_placeholders(field, value):
+    assert not is_placeholder_value(field, value)
+
+
+# ── (c) prompt 硬约束 ──
+
+def test_prompts_carry_hard_rule_and_small_talk_counter_example():
+    from hy_memory.agent import extractor, prompts_zh
+
+    en_rules = [v for v in vars(extractor).values()
+                if isinstance(v, str) and "HARD RULE: fill a field ONLY" in v]
+    zh_rules = [v for v in vars(prompts_zh).values()
+                if isinstance(v, str) and "**硬性规则**" in v]
+    assert len(en_rules) >= 2 and len(zh_rules) >= 2
+    assert "small talk has no basic_info" in EXTRACT_FEW_SHOT_EN
+    assert "寒暄没有 basic_info" in EXTRACT_FEW_SHOT_ZH