[AC-MEMORY-SUMMARY] feat(mid): 新增对话记忆摘要生成器

- 新增 MemorySummaryGenerator 用于生成对话摘要
- 新增 memory_summary_prompt 提供摘要生成提示词模板
- 支持将长对话历史压缩为简洁摘要
- 更新 mid 服务模块导出

ai-service/app/services/mid/__init__.py

@@ -14,6 +14,7 @@ from .metrics_collector import MetricsCollector, SessionMetrics, AggregatedMetri
 from .tool_registry import ToolRegistry, ToolDefinition, ToolExecutionResult, get_tool_registry, init_tool_registry
 from .tool_call_recorder import ToolCallRecorder, ToolCallStatistics, get_tool_call_recorder
 from .memory_adapter import MemoryAdapter, UserMemory
+from .memory_summary_generator import MemorySummaryGenerator
 from .default_kb_tool_runner import DefaultKbToolRunner, KbToolResult, KbToolConfig, get_default_kb_tool_runner
 from .segment_humanizer import SegmentHumanizer, HumanizeConfig, LengthBucket, get_segment_humanizer
 from .runtime_observer import RuntimeObserver, RuntimeContext, get_runtime_observer
@@ -55,6 +56,7 @@ __all__ = [
     "get_tool_call_recorder",
     "MemoryAdapter",
     "UserMemory",
+    "MemorySummaryGenerator",
     "DefaultKbToolRunner",
     "KbToolResult",
     "KbToolConfig",

ai-service/app/services/mid/memory_summary_generator.py

@@ -0,0 +1,58 @@
+"""
+Memory summary generator using LLM.
+[AC-IDMP-14] Generate rolling summary for memory update.
+"""
+
+from __future__ import annotations
+
+import logging
+from typing import Any
+
+from app.services.llm.base import LLMConfig
+from app.services.llm.factory import get_llm_config_manager
+from app.services.mid.memory_summary_prompt import build_memory_summary_prompt
+
+logger = logging.getLogger(__name__)
+
+
+class MemorySummaryGenerator:
+    """
+    LLM-based memory summary generator.
+
+    Output expected to be a JSON object or structured text.
+    """
+
+    def __init__(self, max_tokens: int = 512, temperature: float = 0.2):
+        self._max_tokens = max_tokens
+        self._temperature = temperature
+
+    async def __call__(
+        self,
+        messages: list[dict[str, Any]],
+        old_summary: str | None = None,
+    ) -> str | None:
+        try:
+            llm_manager = get_llm_config_manager()
+            llm_client = llm_manager.get_client()
+        except Exception as e:
+            logger.warning(f"[AC-IDMP-14] Failed to get LLM client: {e}")
+            return None
+
+        prompt = build_memory_summary_prompt(messages, old_summary)
+
+        try:
+            response = await llm_client.generate(
+                messages=[
+                    {"role": "system", "content": "你是一个严格遵循 JSON 格式的摘要器。仅输出 JSON。"},
+                    {"role": "user", "content": prompt},
+                ],
+                config=LLMConfig(
+                    max_tokens=self._max_tokens,
+                    temperature=self._temperature,
+                ),
+            )
+        except Exception as e:
+            logger.warning(f"[AC-IDMP-14] Summary generation failed: {e}")
+            return None
+
+        return response.content

ai-service/app/services/mid/memory_summary_prompt.py

@@ -0,0 +1,46 @@
+"""
+Memory summary prompt builder.
+[AC-IDMP-14] Rolling summary prompt for memory update.
+"""
+
+from __future__ import annotations
+
+from typing import Any
+
+SUMMARY_PROMPT_TEMPLATE = """
+你是一个记忆摘要生成器。你的目标是把对话中稳定、有长期价值的信息归纳为“可用于记忆召回”的摘要。
+要求：
+1) 必须保留：稳定事实、用户偏好、未解决问题；
+2) 不要写闲聊/情绪；
+3) 输出必须为严格 JSON 对象，只允许以下字段：summary, facts, preferences, open_issues；
+4) summary 为一段话（150-300字以内）；facts/preferences/open_issues 为列表；
+5) 如果新内容没有变化，保留旧摘要并可轻微精简；
+6) 所有内容必须基于对话，不允许编造。
+
+【旧摘要】
+{old_summary}
+
+【最近对话】
+{recent_messages}
+""".strip()
+
+
+def build_recent_messages_text(messages: list[dict[str, Any]]) -> str:
+    lines: list[str] = []
+    for msg in messages:
+        role = msg.get("role", "unknown")
+        content = msg.get("content", "")
+        if not content:
+            continue
+        lines.append(f"{role}: {content}")
+    return "\n".join(lines)
+
+
+def build_memory_summary_prompt(
+    messages: list[dict[str, Any]],
+    old_summary: str | None = None,
+) -> str:
+    return SUMMARY_PROMPT_TEMPLATE.format(
+        old_summary=old_summary or "无",
+        recent_messages=build_recent_messages_text(messages),
+    )