记忆系统，上下文处理+mem0

art-agent/backend/app/agent/loop.py

@@ -1,6 +1,8 @@
 """Agent Loop 主循环：LLM 对话 -> 工具调用 -> 结果回传 -> 继续。"""
 
+import asyncio
 import json
+import logging
 import re
 import os
 from typing import AsyncGenerator, Optional
@@ -8,9 +10,17 @@ from typing import AsyncGenerator, Optional
 from openai import AsyncOpenAI
 
 from app.agent.tools import TOOL_DEFINITIONS, execute_tool
-from app.config import get_llm_model, get_llm_max_iterations, get_image_model_config
+from app.config import (
+    get_llm_model,
+    get_llm_max_iterations,
+    get_image_model_config,
+    get_max_recent_turns,
+)
+from app.memory import get_memory
 from app.services.image_gen import to_data_uri
 
+logger = logging.getLogger(__name__)
+
 
 def _get_client() -> AsyncOpenAI:
     base_url = os.getenv("OPENAI_BASE_URL")
@@ -59,6 +69,7 @@ async def run_agent_loop(
     messages: list[dict],
     ref_image_url: Optional[str] = None,
     image_model: Optional[str] = None,
+    session_id: Optional[str] = None,
 ) -> AsyncGenerator[dict, None]:
     """
     运行 Agent Loop，以 SSE 事件流形式 yield 结果。
@@ -67,10 +78,42 @@ async def run_agent_loop(
     - text_delta: LLM 文字增量
     - tool_start: 开始执行工具
     - image_result: 图片生成结果
+    - memory_warning: 记忆存储异常（不中断对话）
     - done: 完成
     - error: 错误
     """
-    # 构建消息列表：将当前生图模型信息注入 system prompt，避免 LLM 猜测模型名
+    # ── 滑动窗口：截断过早的消息 ──
+    max_turns = get_max_recent_turns()
+    if len(messages) > max_turns:
+        messages = messages[-max_turns:]
+
+    # ── 记忆检索：用最新 user 消息语义检索相关记忆 ──
+    memory_block = ""
+    try:
+        memory = get_memory()
+        last_user_content = ""
+        for m in reversed(messages):
+            if m["role"] == "user":
+                last_user_content = m["content"]
+                break
+
+        if last_user_content:
+            search_kwargs = {"query": last_user_content, "user_id": "default_user", "limit": 10}
+            if session_id:
+                search_kwargs["run_id"] = session_id
+            relevant = memory.search(**search_kwargs)
+
+            results = relevant.get("results", []) if isinstance(relevant, dict) else relevant
+            if results:
+                items = "\n".join(f"- {m['memory']}" for m in results if m.get("memory"))
+                if items:
+                    memory_block = f"\n\n## 用户记忆（来自历史对话）\n{items}"
+    except Exception as e:
+        logger.error("Mem0 记忆检索失败: %s", e, exc_info=True)
+        yield {"type": "error", "data": {"message": f"记忆系统检索失败: {e}"}}
+        return
+
+    # ── 构建 system prompt ──
     model_config = get_image_model_config(image_model)
     current_model_name = model_config.get('name', '未知')
     current_model_id = model_config.get('id', '未知')
@@ -81,7 +124,7 @@ async def run_agent_loop(
         f"**注意**：对话历史中可能包含之前使用其他模型的记录，忽略那些旧模型名。"
         f"本次生成使用的是 {current_model_name}，在回复中只能使用这个名称。"
     )
-    api_messages = [{"role": "system", "content": SYSTEM_PROMPT + model_hint}]
+    api_messages = [{"role": "system", "content": SYSTEM_PROMPT + model_hint + memory_block}]
 
     # 检测当前 LLM 是否支持 vision（多模态图片输入）
     llm_model = get_llm_model().lower()
@@ -90,7 +133,6 @@ async def run_agent_loop(
     for msg in messages:
         if msg["role"] == "user" and ref_image_url and msg is messages[-1]:
             if vision_capable:
-                # 支持 vision 的模型：直接发送图片
                 api_messages.append({
                     "role": "user",
                     "content": [
@@ -102,7 +144,6 @@ async def run_agent_loop(
                     ],
                 })
             else:
-                # 不支持 vision 的模型：以文字提示告知有参考图，风格由 IP-Adapter 处理
                 hint = (
                     f"{msg['content']}\n\n"
                     "【系统提示：用户上传了一张参考图，已自动传递给图片生成工具的 IP-Adapter。"
@@ -181,7 +222,9 @@ async def run_agent_loop(
                     ),
                 })
                 continue
-            yield {"type": "done", "data": {}}
+            # 对话正常结束，异步存储记忆
+            async for evt in _store_and_done(messages, session_id):
+                yield evt
             return
 
         # 将 assistant 消息（含 tool_calls）加入历史
@@ -253,6 +296,34 @@ async def run_agent_loop(
         # ref_image_url 不清除：消息构建只在循环外执行一次，
         # 后续工具调用仍需参考图（InstantStyle 等模型必须有 style_image）
 
+    # 迭代次数用尽，存储记忆后结束
+    async for evt in _store_and_done(messages, session_id):
+        yield evt
+
+
+# ─── 异步记忆存储 ─────────────────────────────────────────
+
+
+async def _store_and_done(
+    messages: list[dict], session_id: Optional[str]
+) -> AsyncGenerator[dict, None]:
+    """触发 Mem0 异步存储后 yield done。存储失败时 yield warning 但不中断。"""
+    try:
+        memory = get_memory()
+        recent = messages[-4:] if len(messages) >= 4 else messages
+        add_kwargs: dict = {"user_id": "default_user"}
+        if session_id:
+            add_kwargs["run_id"] = session_id
+
+        loop = asyncio.get_running_loop()
+        await loop.run_in_executor(None, lambda: memory.add(recent, **add_kwargs))
+    except Exception as e:
+        logger.error("Mem0 记忆存储失败: %s", e, exc_info=True)
+        yield {
+            "type": "memory_warning",
+            "data": {"message": f"记忆存储失败: {e}"},
+        }
+
     yield {"type": "done", "data": {}}