ChatAgent之Memory Layer设计

{
	user_id_xxx:[
			{session_id_xxx1:[{
				"role": "user",
				"content": "hi, 我是Louis.很高兴你能看到这篇文章"
				},{
				"role": "assistant",
				"content": "hi Louis, xxxxxxx"
				},
			...
			]},
			{session_id_xxx2:[{
				"role": "user",
				"content": "你好,今天的东京天气怎么样?"
				},{
				"role": "assistant",
				"content": "东京今天的天气阳光明媚,气温15度左右,祝你度过美好的今天.xxxx"
				},
			...
			]},
		...
	],
	user_id_yyy:[...],
	...
}

{
	user_id_xxx:[
			{session_id_xxx1:[{
				"role": "system",
				"content": "用户Louis先是自我介绍了,然后询问了东京今天的天气情况,LLM回复为:东京今天的天气阳光明媚,气温15度左右.xxxx"
				}
			]}
	],
	user_id_yyy:[...],
	...
}

{
  "memory_id": "user_id_xxx_mem_002",
  "tenant_id": "app", 
  "user_id": "user_id_xxx",
  "agent_id": "worker_agent",

  "type": "preference",
  "scope": "private",

  "content": {
    "raw": "请用简洁,幽默的语气来和我聊天。",
    "summary": "User prefers concise and humorous tone in conversations",
    "structured": {
      "tone": ["concise", "humorous"],
      "communication_style": "casual",
      "priority": "high"
    }
  },

  "embedding": {
    "model": "e5-large",
    "vector": [0.123, 0.456, ...]
  },

  "metadata": {
    "importance": 0.92,
    "confidence": 0.96,
    "tags": ["preference", "tone", "style"],
    "source": "explicit_user_input",
    "version": 1
  },

  "time": {
    "created_at": "2026-04-20T10:00:00Z",
    "updated_at": "2026-04-25T08:00:00Z",
    "last_accessed": "2026-04-25T08:00:00Z",
    "decay_score": 0.95
  },

  "relations": [
    {
      "type": "related_to",
      "target_memory_id": "user_id_xxx_mem_001"
    }
  ],

  "access_control": {
    "visibility": "private",
    "allowed_users": [user_id_xxx],
    "allowed_roles": ["admin", "editor"]
  }
}

# 用户的10-20轮会话提取出来的长期记忆,标记为mem_002
"memory_id": "user_id_xxx_mem_002",

# 用户用的是iOS app
"tenant_id": "app_iOS", 

# 用户的ID
"user_id": "user_id_xxx",

# 用户用的ChatAgent名字是“worker_agent”
"agent_id": "worker_agent",

# 长期记忆的类型是 用户偏好. episodic | semantic | procedural | preference | profile
"type": "preference",

# 长期记忆是私有的. private | share | public
"scope": "private",

"content": {

	# 用户query的原始内容
    "raw": "请用简洁,幽默的语气来和我聊天。",

	# LLM总结提取后的长期记忆
    "summary": "User prefers concise and humorous tone in conversations",

	# LLM总结提取后的长期记忆的结构化json
    "structured": {
      "tone": ["concise", "humorous"],
      "communication_style": "casual",
      "priority": "high"
    }
  }

"embedding": {
	# 长期记忆用了微软的"e5-large"模型来embedding.
    "model": "e5-large",
	
	# 向量化后的长期记忆
    "vector": [0.123, 0.456, ...]
  },

"metadata": {
	# 这条长期记忆的重要程度
    "importance": 0.92,

	# 这条长期记忆的可信度
    "confidence": 0.96,
	
	# 这条长期记忆的标签
    "tags": ["preference", "tone", "style"],

	# 这条长期记忆的来源: explicit_user_input | implicit_inference | system_generated | external_data | corrected_memory      

    "source": "explicit_user_input",

	# 这条长期记忆的版本号
    "version": 1
  },

"time": {
	# 这条长期记忆的创建日期
    "created_at": "2026-04-20T10:00:00Z",
	
	# 这条长期记忆的更新日期
    "updated_at": "2026-04-25T08:00:00Z",

	# 这条长期记忆的最新使用时间
    "last_accessed": "2026-04-25T08:00:00Z",

	# 这条长期记忆的衰减分数0-1. 0.95代表着这条长期记忆值得长时间保存,几乎不会被删除.
    "decay_score": 0.95
  },

"relations": [
    {

	# 这条长期记忆是用户的10-20轮会话提取出来的, 上一条记忆mem_001,代表用户0-10轮的会话的长期记忆.
      "type": "related_to",
      "target_memory_id": "user_id_xxx_mem_001"
    }
  ],

# 访问控制
"access_control": {
	# 这条长期记忆是私有的 
    "visibility": "private",

	# 这条长期记忆只允许user_id_xxx访问
    "allowed_users": [user_id_xxx],

	# 这条长期记忆允许"admin", "editor"访问
    "allowed_roles": ["admin", "editor"]
  }