feat:写入跟用户相同的语言的记忆

backend/packages/harness/deerflow/agents/memory/thread_prompt.py

@@ -3,6 +3,7 @@
 from __future__ import annotations
 
 import json
+import re
 from typing import Any
 
 from deerflow.agents.memory.prompt import _coerce_confidence, _count_tokens, format_conversation_for_update
@@ -19,6 +20,8 @@ Conversation:
 {conversation}
 </conversation>
 
+Preferred memory language: {preferred_language}
+
 Return JSON only with this schema:
 {{
   "user": {{
@@ -44,6 +47,7 @@ Rules:
 - Keep only stable and useful user profile facts.
 - Do not store sensitive personal data (phone/email/address/password/token/id/bank).
 - Deduplicate and keep high-confidence facts.
+- Write all human-readable text fields (`summary`, `content`, and similar prose) in the preferred memory language.
 - Return valid JSON only.
 """
 
@@ -64,6 +68,37 @@ def create_empty_thread_memory() -> dict[str, Any]:
     }
 
 
+def _infer_preferred_memory_language(messages: list[Any]) -> str:
+    conversation = format_conversation_for_update(messages)
+    if not conversation.strip():
+        return "same as the user's latest message"
+
+    # If user explicitly provides locale hints, prefer them.
+    locale_match = re.search(r"\b([a-z]{2}-[A-Z]{2})\b", conversation)
+    if locale_match:
+        return locale_match.group(1)
+
+    # Script-based heuristic (dynamic, not hard-coded to two languages).
+    script_patterns = {
+        "zh-Hans": r"[\u4e00-\u9fff]",
+        "ja-JP": r"[\u3040-\u30ff]",
+        "ko-KR": r"[\uac00-\ud7af]",
+        "ru-RU": r"[\u0400-\u04FF]",
+        "ar": r"[\u0600-\u06FF]",
+        "hi-IN": r"[\u0900-\u097F]",
+        "th-TH": r"[\u0E00-\u0E7F]",
+        "he-IL": r"[\u0590-\u05FF]",
+        "el-GR": r"[\u0370-\u03FF]",
+    }
+    counts = {lang: len(re.findall(pattern, conversation)) for lang, pattern in script_patterns.items()}
+    best_lang, best_count = max(counts.items(), key=lambda item: item[1])
+    if best_count > 0:
+        return best_lang
+
+    # Latin-script fallback: ask model to keep same language as the user's latest message.
+    return "same as the user's latest message"
+
+
 def format_thread_memory_for_injection(memory_data: dict[str, Any], max_tokens: int = 2000) -> str:
     if not memory_data:
         return ""
@@ -130,4 +165,5 @@ def build_thread_memory_prompt(existing_memory: dict[str, Any], messages: list[A
     return THREAD_MEMORY_UPDATE_PROMPT.format(
         existing_memory=json.dumps(existing_memory, ensure_ascii=False, indent=2),
         conversation=format_conversation_for_update(messages),
+        preferred_language=_infer_preferred_memory_language(messages),
     )

backend/tests/test_thread_memory_prompt.py

@@ -34,3 +34,20 @@ def test_build_thread_memory_prompt_does_not_raise_format_key_error():
     )
     assert "Current per-thread memory" in prompt
     assert '"user"' in prompt
+    assert "Preferred memory language: same as the user's latest message" in prompt
+
+
+def test_build_thread_memory_prompt_prefers_chinese_for_chinese_conversation():
+    prompt = build_thread_memory_prompt(
+        {"user": {}, "history": {}, "facts": []},
+        [HumanMessage(content="我叫小明，我更喜欢中文交流。")],
+    )
+    assert "Preferred memory language: zh-Hans" in prompt
+
+
+def test_build_thread_memory_prompt_prefers_japanese_for_japanese_conversation():
+    prompt = build_thread_memory_prompt(
+        {"user": {}, "history": {}, "facts": []},
+        [HumanMessage(content="私は日本語で会話したいです。")],
+    )
+    assert "Preferred memory language: ja-JP" in prompt