Generative agents 代码分析 记忆

ai2things2025-07-23 17:23

记忆的基本概念

Recency

最近访问过的记忆对象将会获得更高的分数，这样一来，刚刚发生的事件或者今天早上的事件很可能仍然在虚拟人的注意力范围内。在我们的实现中，我们将Recency视为一个指数衰减函数，根据上次检索记忆以来的沙盒游戏小时数来计算。在本文中衰减因子为0.995。

Importance

给虚拟人认为重要的记忆对象分配更高的分数来区分平凡的记忆和核心记忆。例如，对于日常时间将得到较低的重要性分数，比如在自己的房间吃早餐。与他人分手这样的事情将得到较高的分数。有许多可能的重要性评分实现方式；我们发现直接询问语言模型输出一个整数分数是有效的。完整的提示如下：

在 1 到 10 的范围内，其中 1 是纯粹平凡的（例如，刷牙、整理床铺），10 是印象深刻的事情（例如，分手、大学录取），对以下一段记忆的可能深刻意义进行评分。
记忆：在超市购买东西

评分： <填写>

对于"整理房间"，这个提示返回整数值 2，对于"向你喜欢的人约会"，返回整数值 8。重要性分数是在创建记忆对象时生成的。

Relevance

Relevance 为与当前情况相关的 memory 对象分配更高的分数。什么是相关的取决于 "Related to what？"的答案，因此我们在查询 memory 上设置相关性。

例如，如果查询是学生正在与同学讨论化学考试要学习什么，则关于他们早餐的记忆对象应该具有低相关性，而关于老师和学校作业的记忆对象应该具有较高的相关性。在我们的实现中，我们使用语言模型来生成每个内存的文本描述的嵌入向量。然后，我们将相关性计算为 memory 的嵌入向量和查询 memory 的嵌入向量之间的余弦相似度。

为了计算最终的检索分数，我们将 Recency（新近度）、Importance（重要性）和 Relevance（相关性）分数标准化到 [0,1] 使用最小-最大缩放的范围。检索函数将所有记忆作为三个元素的加权组合进行评分：

。在我们的实现中，所有 α s 都设置为 1。适合语言模型上下文窗口的排名靠前的记忆将包含在提示中。

各种类型的记忆

空间记忆 spatia memory

json 复制代码 
{
  "the Ville": {
    "Hobbs Cafe": {
      "cafe": [
        "refrigerator",
        "cafe customer seating",
        "cooking area",
        "kitchen sink",
        "behind the cafe counter",
        "piano"
      ]
    },
    "Isabella Rodriguez's apartment": {
      "main room": [
        "bed",
        "desk",
        "refrigerator",
        "closet",
        "shelf"
      ],
      "bathroom": [
        "shower",
        "bathroom sink",
        "toilet"
      ]
    },
    "The Rose and Crown Pub": {
      "pub": [
        "shelf",
        "refrigerator",
        "bar customer seating",
        "behind the bar counter",
        "kitchen sink",
        "cooking area",
        "microphone"
      ]
    },
    "Carlos Gomez's apartment": {
      "main room": [
        "desk",
        "bed"
      ],
      "bathroom": [

      ]
    }
  }
}

联想记忆 associative memory

python 复制代码 
# 向量记忆存储
elf.embeddings = json.load(open(f_saved + "/embeddings.json"))
# 记忆节点
nodes_load = json.load(open(f_saved + "/nodes.json"))
# 记录 event 和 thought 的强度（时间长度表示？）
kw_strength_load = json.load(open(f_saved + "/kw_strength.json"))
    if kw_strength_load["kw_strength_event"]: 
      self.kw_strength_event = kw_strength_load["kw_strength_event"]
    if kw_strength_load["kw_strength_thought"]: 
      self.kw_strength_thought = kw_strength_load["kw_strength_thought"]

短期记忆 scratch

json 复制代码 
{
  "vision_r": 8,
  "att_bandwidth": 8,
  "retention": 8,
  "curr_time": "February 14, 2023, 00:02:20",
  "curr_tile": [
    73,
    14
  ],
  "daily_plan_req": "1. Wake up at 6am and complete morning routine (shower, brush teeth, get dressed) by 6:30am. 2. Arrive at Hobbs Cafe by 7:30am to do final preparations for the Valentine's Day party. 3. Welcome and greet customers as they arrive for the party starting at 5pm. 4. Ensure that food and drinks are well-stocked and available throughout the evening. 5. Continuously check on guests and provide excellent customer service during the event. 6. Close the cafe at 8pm and do a final walkthrough to ensure everything is in order before heading home.",
  "name": "Isabella Rodriguez",
  "first_name": "Isabella",
  "last_name": "Rodriguez",
  "age": 34,
  "innate": "friendly, outgoing, hospitable",
  "learned": "Isabella Rodriguez is a cafe owner of Hobbs Cafe who loves to make people feel welcome. She is always looking for ways to make the cafe a place where people can come to relax and enjoy themselves.",
  "currently": "Status: Today is the day! Isabella Rodriguez is filled with excitement as the Valentine's Day party at Hobbs Cafe is finally here. She has been tirelessly planning and preparing for this event, and now it's time to bring it to life. The party is set to start at 5pm at Hobbs Cafe, and Isabella has ensured that everything is carefully arranged and ready to create a memorable evening for everyone. She can't wait to welcome her customers and friends to join the festivities and celebrate love together. Isabella is confident that her efforts will pay off and that the party will be a great success.",
  "lifestyle": "Isabella Rodriguez goes to bed around 11pm, awakes up around 6am.",
  "living_area": "the Ville:Isabella Rodriguez's apartment:main room",
  "concept_forget": 100,
  "daily_reflection_time": 180,
  "daily_reflection_size": 5,
  "overlap_reflect_th": 4,
  "kw_strg_event_reflect_th": 10,
  "kw_strg_thought_reflect_th": 9,
  "recency_w": 1,
  "relevance_w": 1,
  "importance_w": 1,
  "recency_decay": 0.995,
  "importance_trigger_max": 150,
  "importance_trigger_curr": 122,
  "importance_ele_n": 20,
  "thought_count": 5,
  "daily_req": [
    "wake up and complete the morning routine at 6:00 am",
    "open Hobbs Cafe at 8:00 am",
    "have lunch with her staff at 12:00 pm",
    "attend to guests at the cafe from 8:00 am to 8:00 pm",
    "take a long walk after closing the cafe at 8:00 pm"
  ],
  "f_daily_schedule": [
    [
      "sleeping",
      360
    ],
    [
      "waking up and completing her morning routine",
      60
    ],
    [
      "arriving at Hobbs Cafe to do final preparations for the Valentine's Day party",
      60
    ],
    [
      "opening Hobbs Cafe and attending to guests",
      60
    ],
    [
      "having lunch with her staff",
      60
    ],
    [
      "attending to guests at the cafe",
      120
    ],
    [
      "having lunch with her staff",
      60
    ],
    [
      "attending to guests at the cafe",
      240
    ],
    [
      "welcoming and greeting customers as they arrive for the Valentine's Day party",
      60
    ],
    [
      "attending to guests at the cafe",
      120
    ],
    [
      "closing Hobbs Cafe and doing a final walkthrough to ensure everything is in order",
      60
    ],
    [
      "taking a long walk",
      120
    ],
    [
      "going to bed",
      60
    ]
  ],
  "f_daily_schedule_hourly_org": [
    [
      "sleeping",
      360
    ],
    [
      "waking up and completing her morning routine",
      60
    ],
    [
      "arriving at Hobbs Cafe to do final preparations for the Valentine's Day party",
      60
    ],
    [
      "opening Hobbs Cafe and attending to guests",
      60
    ],
    [
      "having lunch with her staff",
      60
    ],
    [
      "attending to guests at the cafe",
      120
    ],
    [
      "having lunch with her staff",
      60
    ],
    [
      "attending to guests at the cafe",
      240
    ],
    [
      "welcoming and greeting customers as they arrive for the Valentine's Day party",
      60
    ],
    [
      "attending to guests at the cafe",
      120
    ],
    [
      "closing Hobbs Cafe and doing a final walkthrough to ensure everything is in order",
      60
    ],
    [
      "taking a long walk",
      120
    ],
    [
      "going to bed",
      60
    ]
  ],
  "act_address": "the Ville:Isabella Rodriguez's apartment:main room:bed",
  "act_start_time": "February 13, 2023, 21:11:00",
  "act_duration": 180,
  "act_description": "sleeping",
  "act_pronunciatio": "\ud83d\udca4",
  "act_event": [
    "Isabella Rodriguez",
    "is",
    "sleep"
  ],
  "act_obj_description": "being used",
  "act_obj_pronunciatio": "\ud83d\udcbb",
  "act_obj_event": [
    "bed",
    "be",
    "used"
  ],
  "chatting_with": null,
  "chat": null,
  "chatting_with_buffer": {
    "Klaus Mueller": -1525,
    "Maria Lopez": -1189
  },
  "chatting_end_time": null,
  "act_path_set": true,
  "planned_path": []
}

记忆存储

记忆节点的数据结构

python 复制代码 
class ConceptNode: 
  def __init__(self,
               node_id, node_count, type_count, node_type, depth,
               created, expiration, 
               s, p, o, 
               description, embedding_key, poignancy, keywords, filling): 
    self.node_id = node_id		# 编号
    self.node_count = node_count
    self.type_count = type_count
    self.type = node_type 	# 类型，event/chat/thought，分别对应 add_event/add_chat/add_thought 方法
    self.depth = depth		# ？？

    self.created = created
    self.expiration = expiration		# 到期时间
    self.last_accessed = self.created

    self.subject = s		# 三元组，主语
    self.predicate = p	# 谓语
    self.object = o			# 宾语

    self.description = description
    self.embedding_key = embedding_key
    self.poignancy = poignancy
    self.keywords = keywords
    self.filling = filling

记忆明文

记忆节点的 description 字段内容会同时转换为向量数据保存在 embeddings.json中，方便后面做检索 Reveriev ：

json 复制代码 
{
  "node_147": {
    "node_count": 147,
    "type_count": 2,
    "type": "chat",
    "depth": 0,
    "created": "2023-02-13 11:22:40",
    "expiration": null,
    "subject": "Klaus Mueller",
    "predicate": "chat with",
    "object": "Isabella Rodriguez",
    "description": "conversing about a conversation about Isabella inviting Klaus to her Valentine's Day party at Hobbs Cafe on February 14th, 2023 from 5pm to 7pm.",
    "embedding_key": "conversing about a conversation about Isabella inviting Klaus to her Valentine's Day party at Hobbs Cafe on February 14th, 2023 from 5pm to 7pm.",
    "poignancy": 6,
    "keywords": ["Klaus Mueller", "Isabella Rodriguez"],
    "filling": [
      [
        "Isabella Rodriguez",
        "Hi Klaus! How are you enjoying your meal? I wanted to let you know that I'm planning a Valentine's Day party at Hobbs Cafe on February 14th, 2023 from 5pm to 7pm. I would love for you to join us!"
      ],
      [
        "Klaus Mueller",
        "Oh, hi Isabella! I'm doing well, thank you. The meal is delicious as always. A Valentine's Day party sounds fun. I'd love to join! Thank you for inviting me."
      ]
    ]
  },
  "node_94": {
    "node_count": 94,
    "type_count": 12,
    "type": "thought",
    "depth": 2,
    "created": "2023-02-13 09:15:10",
    "expiration": "2023-03-15 09:15:10",
    "subject": "Klaus Mueller",
    "predicate": "have",
    "object": "strong relationship",
    "description": "Klaus Mueller and Maria Lopez have a strong relationship",
    "embedding_key": "Klaus Mueller and Maria Lopez have a strong relationship",
    "poignancy": 7,
    "keywords": ["have", "Klaus Mueller", "strong relationship"],
    "filling": ["node_10", "node_11", "node_9"]
  },
  <!-- -->
  "node_11": {
    "node_count": 11,
    "type_count": 4,
    "type": "thought",
    "depth": 1,
    "created": "2023-02-13 00:00:20",
    "expiration": "2023-03-15 00:00:20",
    "subject": "Klaus Mueller",
    "predicate": "be",
    "object": "close friends and classmates",
    "description": "Klaus Mueller is close friends and classmates with Maria Lopez.",
    "embedding_key": "Klaus Mueller is close friends and classmates with Maria Lopez.",
    "poignancy": 6,
    "keywords": ["Klaus Mueller", "be", "close friends and classmates"],
    "filling": null
  },
  "node_10": {
    "node_count": 10,
    "type_count": 3,
    "type": "thought",
    "depth": 1,
    "created": "2023-02-13 00:00:20",
    "expiration": "2023-03-15 00:00:20",
    "subject": "Klaus Mueller",
    "predicate": "have been",
    "object": "friends",
    "description": "Klaus Mueller and Maria Lopez have been friends for more than 2 years.",
    "embedding_key": "Klaus Mueller and Maria Lopez have been friends for more than 2 years.",
    "poignancy": 7,
    "keywords": ["friends", "Klaus Mueller", "have been"],
    "filling": null
  },
  "node_9": {
    "node_count": 9,
    "type_count": 2,
    "type": "thought",
    "depth": 1,
    "created": "2023-02-13 00:00:20",
    "expiration": "2023-03-15 00:00:20",
    "subject": "Klaus Mueller",
    "predicate": "have",
    "object": "crush on Maria Lopez",
    "description": "Klaus Mueller has a crush on Maria Lopez.",
    "embedding_key": "Klaus Mueller has a crush on Maria Lopez.",
    "poignancy": 6,
    "keywords": ["have", "Klaus Mueller", "crush on Maria Lopez"],
    "filling": null
  },
}

记忆向量化

key + 向量值，其中 key 对应 node.embedding_key

json 复制代码 
{
    "Isabella Rodriguez is idle": [
        -0.03697006031870842,
        -0.010967422276735306,
        0.0015322132967412472,
        -0.00994198489934206,
        -0.014587010256946087,
        0.002556801540777087,
    ]
}

自我认知（角色定义）

所有记忆最终都是赋予一个具体 persona 的。

python 复制代码 
# 构造角色，如果有角色相关的本地记忆，则加载记忆，没有就创建新的记忆

# environment/frontend_server/storage/July1_the_ville_isabella_maria_klaus-step-3-20/personas/Isabella Rodriguez/bootstrap_memory

# 空间记忆
f_s_mem_saved = f"{folder_mem_saved}/bootstrap_memory/spatial_memory.json"
self.s_mem = MemoryTree(f_s_mem_saved)

# 联想记忆
f_a_mem_saved = f"{folder_mem_saved}/bootstrap_memory/associative_memory"
    self.a_mem = AssociativeMemory(f_a_mem_saved)

# 短期记忆，包含了个人信息
scratch_saved = f"{folder_mem_saved}/bootstrap_memory/scratch.json"
self.scratch = Scratch(scratch_saved)
相关推荐
c76914 小时前
【文献笔记】ARS: Automatic Routing Solver with Large Language Models
人工智能·笔记·语言模型·自然语言处理·llm·论文笔记·cvrp
程序员X小鹿17 小时前
大厂出手！用了这个更厉害的P图AI，我把PS卸了！文字也能无痕修改！1秒救废图（附保姆级教程）
agent
大模型开发20 小时前
刷到就是赚到！大模型学习经验分享，帮你少走 3 年弯路
程序员·llm·deepseek
AI大模型21 小时前
五分钟搭建属于你的AI助手：Ollama+DeepSeek+AnythingLLM深度整合
llm·ollama·deepseek
AI大模型21 小时前
关于智能体（AI Agent），不得不看的一篇总结（建议收藏）
程序员·llm·agent
Code_流苏21 小时前
ChatGPT Agent深度解析：告别单纯问答，一个指令搞定复杂任务？
人工智能·自然语言处理·chatgpt·openai·agent·智能体
聚客AI1 天前
LLM→RAG→Agent→Training的企业级AI应用落地分层实施路线图
人工智能·llm·agent
雷渊1 天前
LangChain4j-ChatMessage的种类以及作用
langchain·llm
热门推荐
01全球最强模型Grok4，国内已可免费使用！（附教程）02【2025.7.18】更新vscode后所有.vue文件template标签后报红的临时解决办法，Vue - Official 插件3.0.2导致03Claude Code用不了？来试下Qwen3-Coder加持的Qwen Code吧04Cursor Claude 模型无法使用的解决方法05这次领先Cursor！体验了Trae 2.0 SOLO 模式，超酷！06KGG转MP3工具|非KGM文件|解密音频07我亲历了2025年最荒谬的前端灾难：一支触控笔"干掉"了全球CSS预处理器08Cursor限制国内用户使用Claude模型？附简单解决方案！09赛博保姆出列！Gemini CLI & Claude Code + Kimi 手把手教程10Claude Code 最新版已经支持 Windows 安装使用！