Losita/mai-bot
0
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

refactor: 全部代码格式化

Browse Source
This commit is contained in:
Rikki
2025-03-30 04:56:46 +08:00
parent 7adaa2f5a8
commit b2fc824afd
21 changed files with 491 additions and 514 deletions
Download Patch File Download Diff File Expand all files Collapse all files

242
src/plugins/memory_system/Hippocampus.py
View File

@@ -11,7 +11,7 @@ from collections import Counter
from ...common.database import db
from ...plugins.models.utils_model import LLM_request
from src.common.logger import get_module_logger, LogConfig, MEMORY_STYLE_CONFIG
from src.plugins.memory_system.sample_distribution import MemoryBuildScheduler #分布生成器
from src.plugins.memory_system.sample_distribution import MemoryBuildScheduler # 分布生成器
from .memory_config import MemoryConfig
@@ -56,6 +56,7 @@ def get_closest_chat_from_db(length: int, timestamp: str):
return []
def calculate_information_content(text):
"""计算文本的信息量（熵）"""
char_count = Counter(text)
@@ -68,6 +69,7 @@ def calculate_information_content(text):
return entropy
def cosine_similarity(v1, v2):
"""计算余弦相似度"""
dot_product = np.dot(v1, v2)
@@ -223,7 +225,8 @@ class Memory_graph:
return None
#负责海马体与其他部分的交互
# 负责海马体与其他部分的交互
class EntorhinalCortex:
def __init__(self, hippocampus):
self.hippocampus = hippocampus
@@ -243,7 +246,7 @@ class EntorhinalCortex:
n_hours2=self.config.memory_build_distribution[3],
std_hours2=self.config.memory_build_distribution[4],
weight2=self.config.memory_build_distribution[5],
total_samples=self.config.build_memory_sample_num
total_samples=self.config.build_memory_sample_num,
)
timestamps = sample_scheduler.get_timestamp_array()
@@ -251,9 +254,7 @@ class EntorhinalCortex:
chat_samples = []
for timestamp in timestamps:
messages = self.random_get_msg_snippet(
timestamp,
self.config.build_memory_sample_length,
max_memorized_time_per_msg
timestamp, self.config.build_memory_sample_length, max_memorized_time_per_msg
)
if messages:
time_diff = (datetime.datetime.now().timestamp() - timestamp) / 3600
@@ -455,25 +456,25 @@ class EntorhinalCortex:
"""清空数据库并重新同步所有记忆数据"""
start_time = time.time()
logger.info("[数据库] 开始重新同步所有记忆数据...")
# 清空数据库
clear_start = time.time()
db.graph_data.nodes.delete_many({})
db.graph_data.edges.delete_many({})
clear_end = time.time()
logger.info(f"[数据库] 清空数据库耗时: {clear_end - clear_start:.2f}秒")
# 获取所有节点和边
memory_nodes = list(self.memory_graph.G.nodes(data=True))
memory_edges = list(self.memory_graph.G.edges(data=True))
# 重新写入节点
node_start = time.time()
for concept, data in memory_nodes:
memory_items = data.get("memory_items", [])
if not isinstance(memory_items, list):
memory_items = [memory_items] if memory_items else []
node_data = {
"concept": concept,
"memory_items": memory_items,
@@ -484,7 +485,7 @@ class EntorhinalCortex:
db.graph_data.nodes.insert_one(node_data)
node_end = time.time()
logger.info(f"[数据库] 写入 {len(memory_nodes)} 个节点耗时: {node_end - node_start:.2f}秒")
# 重新写入边
edge_start = time.time()
for source, target, data in memory_edges:
@@ -499,12 +500,13 @@ class EntorhinalCortex:
db.graph_data.edges.insert_one(edge_data)
edge_end = time.time()
logger.info(f"[数据库] 写入 {len(memory_edges)} 条边耗时: {edge_end - edge_start:.2f}秒")
end_time = time.time()
logger.success(f"[数据库] 重新同步完成，总耗时: {end_time - start_time:.2f}秒")
logger.success(f"[数据库] 同步了 {len(memory_nodes)} 个节点和 {len(memory_edges)} 条边")
#负责整合，遗忘，合并记忆
# 负责整合，遗忘，合并记忆
class ParahippocampalGyrus:
def __init__(self, hippocampus):
self.hippocampus = hippocampus
@@ -567,26 +569,26 @@ class ParahippocampalGyrus:
topic_num = self.hippocampus.calculate_topic_num(input_text, compress_rate)
topics_response = await self.hippocampus.llm_topic_judge.generate_response(
self.hippocampus.find_topic_llm(input_text, topic_num))
self.hippocampus.find_topic_llm(input_text, topic_num)
)
# 使用正则表达式提取<>中的内容
topics = re.findall(r'<([^>]+)>', topics_response[0])
topics = re.findall(r"<([^>]+)>", topics_response[0])
# 如果没有找到<>包裹的内容，返回['none']
if not topics:
topics = ['none']
topics = ["none"]
else:
# 处理提取出的话题
topics = [
topic.strip()
for topic in ','.join(topics).replace("，", ",").replace("、", ",").replace(" ", ",").split(",")
for topic in ",".join(topics).replace("，", ",").replace("、", ",").replace(" ", ",").split(",")
if topic.strip()
]
# 过滤掉包含禁用关键词的topic
filtered_topics = [
topic for topic in topics
if not any(keyword in topic for keyword in self.config.memory_ban_words)
topic for topic in topics if not any(keyword in topic for keyword in self.config.memory_ban_words)
]
logger.debug(f"过滤后话题: {filtered_topics}")
@@ -601,12 +603,12 @@ class ParahippocampalGyrus:
# 等待所有任务完成
compressed_memory = set()
similar_topics_dict = {}
for topic, task in tasks:
response = await task
if response:
compressed_memory.add((topic, response[0]))
existing_topics = list(self.memory_graph.G.nodes())
similar_topics = []
@@ -651,7 +653,7 @@ class ParahippocampalGyrus:
current_time = datetime.datetime.now().timestamp()
logger.debug(f"添加节点: {', '.join(topic for topic, _ in compressed_memory)}")
all_added_nodes.extend(topic for topic, _ in compressed_memory)
for topic, memory in compressed_memory:
self.memory_graph.add_dot(topic, memory)
all_topics.append(topic)
@@ -661,13 +663,13 @@ class ParahippocampalGyrus:
for similar_topic, similarity in similar_topics:
if topic != similar_topic:
strength = int(similarity * 10)
logger.debug(f"连接相似节点: {topic} 和 {similar_topic} (强度: {strength})")
all_added_edges.append(f"{topic}-{similar_topic}")
all_connected_nodes.append(topic)
all_connected_nodes.append(similar_topic)
self.memory_graph.G.add_edge(
topic,
similar_topic,
@@ -685,14 +687,11 @@ class ParahippocampalGyrus:
logger.success(f"更新记忆: {', '.join(all_added_nodes)}")
logger.debug(f"强化连接: {', '.join(all_added_edges)}")
logger.info(f"强化连接节点: {', '.join(all_connected_nodes)}")
await self.hippocampus.entorhinal_cortex.sync_memory_to_db()
end_time = time.time()
logger.success(
f"---------------------记忆构建耗时: {end_time - start_time:.2f} "
"秒---------------------"
)
logger.success(f"---------------------记忆构建耗时: {end_time - start_time:.2f} 秒---------------------")
async def operation_forget_topic(self, percentage=0.005):
start_time = time.time()
@@ -714,11 +713,11 @@ class ParahippocampalGyrus:
# 使用列表存储变化信息
edge_changes = {
"weakened": [], # 存储减弱的边
"removed": [] # 存储移除的边
"removed": [], # 存储移除的边
}
node_changes = {
"reduced": [], # 存储减少记忆的节点
"removed": [] # 存储移除的节点
"reduced": [], # 存储减少记忆的节点
"removed": [], # 存储移除的节点
}
current_time = datetime.datetime.now().timestamp()
@@ -771,35 +770,40 @@ class ParahippocampalGyrus:
if any(edge_changes.values()) or any(node_changes.values()):
sync_start = time.time()
await self.hippocampus.entorhinal_cortex.resync_memory_to_db()
sync_end = time.time()
logger.info(f"[遗忘] 数据库同步耗时: {sync_end - sync_start:.2f}秒")
# 汇总输出所有变化
logger.info("[遗忘] 遗忘操作统计:")
if edge_changes["weakened"]:
logger.info(
f"[遗忘] 减弱的连接 ({len(edge_changes['weakened'])}个): {', '.join(edge_changes['weakened'])}")
f"[遗忘] 减弱的连接 ({len(edge_changes['weakened'])}个): {', '.join(edge_changes['weakened'])}"
)
if edge_changes["removed"]:
logger.info(
f"[遗忘] 移除的连接 ({len(edge_changes['removed'])}个): {', '.join(edge_changes['removed'])}")
f"[遗忘] 移除的连接 ({len(edge_changes['removed'])}个): {', '.join(edge_changes['removed'])}"
)
if node_changes["reduced"]:
logger.info(
f"[遗忘] 减少记忆的节点 ({len(node_changes['reduced'])}个): {', '.join(node_changes['reduced'])}")
f"[遗忘] 减少记忆的节点 ({len(node_changes['reduced'])}个): {', '.join(node_changes['reduced'])}"
)
if node_changes["removed"]:
logger.info(
f"[遗忘] 移除的节点 ({len(node_changes['removed'])}个): {', '.join(node_changes['removed'])}")
f"[遗忘] 移除的节点 ({len(node_changes['removed'])}个): {', '.join(node_changes['removed'])}"
)
else:
logger.info("[遗忘] 本次检查没有节点或连接满足遗忘条件")
end_time = time.time()
logger.info(f"[遗忘] 总耗时: {end_time - start_time:.2f}秒")
# 海马体
class Hippocampus:
def __init__(self):
@@ -817,8 +821,8 @@ class Hippocampus:
self.parahippocampal_gyrus = ParahippocampalGyrus(self)
# 从数据库加载记忆图
self.entorhinal_cortex.sync_memory_from_db()
self.llm_topic_judge = LLM_request(self.config.llm_topic_judge,request_type="memory")
self.llm_summary_by_topic = LLM_request(self.config.llm_summary_by_topic,request_type="memory")
self.llm_topic_judge = LLM_request(self.config.llm_topic_judge, request_type="memory")
self.llm_summary_by_topic = LLM_request(self.config.llm_summary_by_topic, request_type="memory")
def get_all_node_names(self) -> list:
"""获取记忆图中所有节点的名字列表"""
@@ -901,16 +905,21 @@ class Hippocampus:
memory_items = node_data.get("memory_items", [])
if not isinstance(memory_items, list):
memory_items = [memory_items] if memory_items else []
memories.append((node, memory_items, similarity))
# 按相似度降序排序
memories.sort(key=lambda x: x[2], reverse=True)
return memories
async def get_memory_from_text(self, text: str, max_memory_num: int = 3, max_memory_length: int = 2,
max_depth: int = 3,
fast_retrieval: bool = False) -> list:
async def get_memory_from_text(
self,
text: str,
max_memory_num: int = 3,
max_memory_length: int = 2,
max_depth: int = 3,
fast_retrieval: bool = False,
) -> list:
"""从文本中提取关键词并获取相关记忆。
Args:
@@ -943,18 +952,16 @@ class Hippocampus:
# 使用LLM提取关键词
topic_num = min(5, max(1, int(len(text) * 0.1))) # 根据文本长度动态调整关键词数量
# logger.info(f"提取关键词数量: {topic_num}")
topics_response = await self.llm_topic_judge.generate_response(
self.find_topic_llm(text, topic_num)
)
topics_response = await self.llm_topic_judge.generate_response(self.find_topic_llm(text, topic_num))
# 提取关键词
keywords = re.findall(r'<([^>]+)>', topics_response[0])
keywords = re.findall(r"<([^>]+)>", topics_response[0])
if not keywords:
keywords = []
else:
keywords = [
keyword.strip()
for keyword in ','.join(keywords).replace("，", ",").replace("、", ",").replace(" ", ",").split(",")
for keyword in ",".join(keywords).replace("，", ",").replace("、", ",").replace(" ", ",").split(",")
if keyword.strip()
]
@@ -965,7 +972,7 @@ class Hippocampus:
if not valid_keywords:
logger.info("没有找到有效的关键词节点")
return []
logger.info(f"有效的关键词: {', '.join(valid_keywords)}")
# 从每个关键词获取记忆
@@ -981,35 +988,36 @@ class Hippocampus:
visited_nodes = {keyword}
# 待处理的节点队列，每个元素是(节点, 激活值, 当前深度)
nodes_to_process = [(keyword, 1.0, 0)]
while nodes_to_process:
current_node, current_activation, current_depth = nodes_to_process.pop(0)
# 如果激活值小于0或超过最大深度，停止扩散
if current_activation <= 0 or current_depth >= max_depth:
continue
# 获取当前节点的所有邻居
neighbors = list(self.memory_graph.G.neighbors(current_node))
for neighbor in neighbors:
if neighbor in visited_nodes:
continue
# 获取连接强度
edge_data = self.memory_graph.G[current_node][neighbor]
strength = edge_data.get("strength", 1)
# 计算新的激活值
new_activation = current_activation - (1 / strength)
if new_activation > 0:
activation_values[neighbor] = new_activation
visited_nodes.add(neighbor)
nodes_to_process.append((neighbor, new_activation, current_depth + 1))
logger.debug(
f"节点 '{neighbor}' 被激活，激活值: {new_activation:.2f} (通过 '{current_node}' 连接，强度: {strength}, 深度: {current_depth + 1})") # noqa: E501
f"节点 '{neighbor}' 被激活，激活值: {new_activation:.2f} (通过 '{current_node}' 连接，强度: {strength}, 深度: {current_depth + 1})"
) # noqa: E501
# 更新激活映射
for node, activation_value in activation_values.items():
if activation_value > 0:
@@ -1017,7 +1025,7 @@ class Hippocampus:
activate_map[node] += activation_value
else:
activate_map[node] = activation_value
# 输出激活映射
# logger.info("激活映射统计:")
# for node, total_activation in sorted(activate_map.items(), key=lambda x: x[1], reverse=True):
@@ -1026,28 +1034,24 @@ class Hippocampus:
# 基于激活值平方的独立概率选择
remember_map = {}
# logger.info("基于激活值平方的归一化选择:")
# 计算所有激活值的平方和
total_squared_activation = sum(activation ** 2 for activation in activate_map.values())
total_squared_activation = sum(activation**2 for activation in activate_map.values())
if total_squared_activation > 0:
# 计算归一化的激活值
normalized_activations = {
node: (activation ** 2) / total_squared_activation
for node, activation in activate_map.items()
node: (activation**2) / total_squared_activation for node, activation in activate_map.items()
}
# 按归一化激活值排序并选择前max_memory_num个
sorted_nodes = sorted(
normalized_activations.items(),
key=lambda x: x[1],
reverse=True
)[:max_memory_num]
sorted_nodes = sorted(normalized_activations.items(), key=lambda x: x[1], reverse=True)[:max_memory_num]
# 将选中的节点添加到remember_map
for node, normalized_activation in sorted_nodes:
remember_map[node] = activate_map[node] # 使用原始激活值
logger.debug(
f"节点 '{node}' (归一化激活值: {normalized_activation:.2f}, 激活值: {activate_map[node]:.2f})")
f"节点 '{node}' (归一化激活值: {normalized_activation:.2f}, 激活值: {activate_map[node]:.2f})"
)
else:
logger.info("没有有效的激活值")
@@ -1060,7 +1064,7 @@ class Hippocampus:
memory_items = node_data.get("memory_items", [])
if not isinstance(memory_items, list):
memory_items = [memory_items] if memory_items else []
if memory_items:
logger.debug(f"节点包含 {len(memory_items)} 条记忆")
# 计算每条记忆与输入文本的相似度
@@ -1079,7 +1083,7 @@ class Hippocampus:
memory_similarities.sort(key=lambda x: x[1], reverse=True)
# 获取最匹配的记忆
top_memories = memory_similarities[:max_memory_length]
# 添加到结果中
for memory, similarity in top_memories:
all_memories.append((node, [memory], similarity))
@@ -1106,11 +1110,10 @@ class Hippocampus:
memory = memory_items[0] # 因为每个topic只有一条记忆
result.append((topic, memory))
logger.info(f"选中记忆: {memory} (来自节点: {topic})")
return result
async def get_activate_from_text(self, text: str, max_depth: int = 3,
fast_retrieval: bool = False) -> float:
async def get_activate_from_text(self, text: str, max_depth: int = 3, fast_retrieval: bool = False) -> float:
"""从文本中提取关键词并获取相关记忆。
Args:
@@ -1140,18 +1143,16 @@ class Hippocampus:
# 使用LLM提取关键词
topic_num = min(5, max(1, int(len(text) * 0.1))) # 根据文本长度动态调整关键词数量
# logger.info(f"提取关键词数量: {topic_num}")
topics_response = await self.llm_topic_judge.generate_response(
self.find_topic_llm(text, topic_num)
)
topics_response = await self.llm_topic_judge.generate_response(self.find_topic_llm(text, topic_num))
# 提取关键词
keywords = re.findall(r'<([^>]+)>', topics_response[0])
keywords = re.findall(r"<([^>]+)>", topics_response[0])
if not keywords:
keywords = []
else:
keywords = [
keyword.strip()
for keyword in ','.join(keywords).replace("，", ",").replace("、", ",").replace(" ", ",").split(",")
for keyword in ",".join(keywords).replace("，", ",").replace("、", ",").replace(" ", ",").split(",")
if keyword.strip()
]
@@ -1162,7 +1163,7 @@ class Hippocampus:
if not valid_keywords:
logger.info("没有找到有效的关键词节点")
return 0
logger.info(f"有效的关键词: {', '.join(valid_keywords)}")
# 从每个关键词获取记忆
@@ -1177,35 +1178,35 @@ class Hippocampus:
visited_nodes = {keyword}
# 待处理的节点队列，每个元素是(节点, 激活值, 当前深度)
nodes_to_process = [(keyword, 1.0, 0)]
while nodes_to_process:
current_node, current_activation, current_depth = nodes_to_process.pop(0)
# 如果激活值小于0或超过最大深度，停止扩散
if current_activation <= 0 or current_depth >= max_depth:
continue
# 获取当前节点的所有邻居
neighbors = list(self.memory_graph.G.neighbors(current_node))
for neighbor in neighbors:
if neighbor in visited_nodes:
continue
# 获取连接强度
edge_data = self.memory_graph.G[current_node][neighbor]
strength = edge_data.get("strength", 1)
# 计算新的激活值
new_activation = current_activation - (1 / strength)
if new_activation > 0:
activation_values[neighbor] = new_activation
visited_nodes.add(neighbor)
nodes_to_process.append((neighbor, new_activation, current_depth + 1))
# logger.debug(
# f"节点 '{neighbor}' 被激活，激活值: {new_activation:.2f} (通过 '{current_node}' 连接，强度: {strength}, 深度: {current_depth + 1})") # noqa: E501
# f"节点 '{neighbor}' 被激活，激活值: {new_activation:.2f} (通过 '{current_node}' 连接，强度: {strength}, 深度: {current_depth + 1})") # noqa: E501
# 更新激活映射
for node, activation_value in activation_values.items():
if activation_value > 0:
@@ -1213,23 +1214,24 @@ class Hippocampus:
activate_map[node] += activation_value
else:
activate_map[node] = activation_value
# 输出激活映射
# logger.info("激活映射统计:")
# for node, total_activation in sorted(activate_map.items(), key=lambda x: x[1], reverse=True):
# logger.info(f"节点 '{node}': 累计激活值 = {total_activation:.2f}")
# 计算激活节点数与总节点数的比值
total_activation = sum(activate_map.values())
logger.info(f"总激活值: {total_activation:.2f}")
total_nodes = len(self.memory_graph.G.nodes())
# activated_nodes = len(activate_map)
activation_ratio = total_activation / total_nodes if total_nodes > 0 else 0
activation_ratio = activation_ratio*60
activation_ratio = activation_ratio * 60
logger.info(f"总激活值: {total_activation:.2f}, 总节点数: {total_nodes}, 激活: {activation_ratio}")
return activation_ratio
class HippocampusManager:
_instance = None
_hippocampus = None
@@ -1252,12 +1254,12 @@ class HippocampusManager:
"""初始化海马体实例"""
if self._initialized:
return self._hippocampus
self._global_config = global_config
self._hippocampus = Hippocampus()
self._hippocampus.initialize(global_config)
self._initialized = True
# 输出记忆系统参数信息
config = self._hippocampus.config
@@ -1265,16 +1267,15 @@ class HippocampusManager:
memory_graph = self._hippocampus.memory_graph.G
node_count = len(memory_graph.nodes())
edge_count = len(memory_graph.edges())
logger.success(f'''--------------------------------
logger.success(f"""--------------------------------
记忆系统参数配置:
构建间隔: {global_config.build_memory_interval}秒|样本数: {config.build_memory_sample_num},长度: {config.build_memory_sample_length}|压缩率: {config.memory_compress_rate}
记忆构建分布: {config.memory_build_distribution}
遗忘间隔: {global_config.forget_memory_interval}秒|遗忘比例: {global_config.memory_forget_percentage}|遗忘: {config.memory_forget_time}小时之后
记忆图统计信息: 节点数量: {node_count}, 连接数量: {edge_count}
--------------------------------''') #noqa: E501
--------------------------------""") # noqa: E501
return self._hippocampus
async def build_memory(self):
@@ -1289,17 +1290,22 @@ class HippocampusManager:
raise RuntimeError("HippocampusManager 尚未初始化，请先调用 initialize 方法")
return await self._hippocampus.parahippocampal_gyrus.operation_forget_topic(percentage)
async def get_memory_from_text(self, text: str, max_memory_num: int = 3,
max_memory_length: int = 2, max_depth: int = 3,
fast_retrieval: bool = False) -> list:
async def get_memory_from_text(
self,
text: str,
max_memory_num: int = 3,
max_memory_length: int = 2,
max_depth: int = 3,
fast_retrieval: bool = False,
) -> list:
"""从文本中获取相关记忆的公共接口"""
if not self._initialized:
raise RuntimeError("HippocampusManager 尚未初始化，请先调用 initialize 方法")
return await self._hippocampus.get_memory_from_text(
text, max_memory_num, max_memory_length, max_depth, fast_retrieval)
text, max_memory_num, max_memory_length, max_depth, fast_retrieval
)
async def get_activate_from_text(self, text: str, max_depth: int = 3,
fast_retrieval: bool = False) -> float:
async def get_activate_from_text(self, text: str, max_depth: int = 3, fast_retrieval: bool = False) -> float:
"""从文本中获取激活值的公共接口"""
if not self._initialized:
raise RuntimeError("HippocampusManager 尚未初始化，请先调用 initialize 方法")
@@ -1316,5 +1322,3 @@ class HippocampusManager:
if not self._initialized:
raise RuntimeError("HippocampusManager 尚未初始化，请先调用 initialize 方法")
return self._hippocampus.get_all_node_names()

23
src/plugins/memory_system/debug_memory.py
View File

@@ -3,11 +3,13 @@ import asyncio
import time
import sys
import os
# 添加项目根目录到系统路径
sys.path.append(os.path.dirname(os.path.dirname(os.path.dirname(os.path.dirname(__file__)))))
from src.plugins.memory_system.Hippocampus import HippocampusManager
from src.plugins.config.config import global_config
async def test_memory_system():
"""测试记忆系统的主要功能"""
try:
@@ -24,7 +26,7 @@ async def test_memory_system():
# 测试记忆检索
test_text = "千石可乐在群里聊天"
test_text = '''[03-24 10:39:37] 麦麦(ta的id:2814567326): 早说散步结果下雨改成室内运动啊
test_text = """[03-24 10:39:37] 麦麦(ta的id:2814567326): 早说散步结果下雨改成室内运动啊
[03-24 10:39:37] 麦麦(ta的id:2814567326): [回复：变量] 变量就像今天计划总变
[03-24 10:39:44] 状态异常(ta的id:535554838): 要把本地文件改成弹出来的路径吗
[03-24 10:40:35] 状态异常(ta的id:535554838): [图片：这张图片显示的是Windows系统的环境变量设置界面。界面左侧列出了多个环境变量的值，包括Intel Dev Redist、Windows、Windows PowerShell、OpenSSH、NVIDIA Corporation的目录等。右侧有新建、编辑、浏览、删除、上移、下移和编辑文本等操作按钮。图片下方有一个错误提示框，显示"Windows找不到文件'mongodb\\bin\\mongod.exe'。请确定文件名是否正确后，再试一次。"这意味着用户试图运行MongoDB的mongod.exe程序时，系统找不到该文件。这可能是因为MongoDB的安装路径未正确添加到系统环境变量中，或者文件路径有误。
@@ -39,28 +41,21 @@ async def test_memory_system():
[03-24 10:46:12] (ta的id:3229291803): [表情包：这张表情包显示了一只手正在做"点赞"的动作，通常表示赞同、喜欢或支持。这个表情包所表达的情感是积极的、赞同的或支持的。]
[03-24 10:46:37] 星野風禾(ta的id:2890165435): 还能思考高达
[03-24 10:46:39] 星野風禾(ta的id:2890165435): 什么知识库
[03-24 10:46:49] ❦幻凌慌てない(ta的id:2459587037): 为什么改了回复系数麦麦还是不怎么回复？大佬们''' # noqa: E501
[03-24 10:46:49] ❦幻凌慌てない(ta的id:2459587037): 为什么改了回复系数麦麦还是不怎么回复？大佬们""" # noqa: E501
# test_text = '''千石可乐：分不清AI的陪伴和人类的陪伴,是这样吗？'''
print(f"开始测试记忆检索，测试文本: {test_text}\n")
memories = await hippocampus_manager.get_memory_from_text(
text=test_text,
max_memory_num=3,
max_memory_length=2,
max_depth=3,
fast_retrieval=False
text=test_text, max_memory_num=3, max_memory_length=2, max_depth=3, fast_retrieval=False
)
await asyncio.sleep(1)
print("检索到的记忆:")
for topic, memory_items in memories:
print(f"主题: {topic}")
print(f"- {memory_items}")
# 测试记忆遗忘
# forget_start_time = time.time()
# # print("开始测试记忆遗忘...")
@@ -80,6 +75,7 @@ async def test_memory_system():
print(f"测试过程中出现错误: {e}")
raise
async def main():
"""主函数"""
try:
@@ -91,5 +87,6 @@ async def main():
print(f"程序执行出错: {e}")
raise
if __name__ == "__main__":
asyncio.run(main())
asyncio.run(main())

12
src/plugins/memory_system/memory_config.py
View File

@@ -1,24 +1,26 @@
from dataclasses import dataclass
from typing import List
@dataclass
class MemoryConfig:
"""记忆系统配置类"""
# 记忆构建相关配置
memory_build_distribution: List[float] # 记忆构建的时间分布参数
build_memory_sample_num: int # 每次构建记忆的样本数量
build_memory_sample_length: int # 每个样本的消息长度
memory_compress_rate: float # 记忆压缩率
# 记忆遗忘相关配置
memory_forget_time: int # 记忆遗忘时间（小时）
# 记忆过滤相关配置
memory_ban_words: List[str] # 记忆过滤词列表
llm_topic_judge: str # 话题判断模型
llm_summary_by_topic: str # 话题总结模型
@classmethod
def from_global_config(cls, global_config):
"""从全局配置创建记忆系统配置"""
@@ -30,5 +32,5 @@ class MemoryConfig:
memory_forget_time=global_config.memory_forget_time,
memory_ban_words=global_config.memory_ban_words,
llm_topic_judge=global_config.llm_topic_judge,
llm_summary_by_topic=global_config.llm_summary_by_topic
)
llm_summary_by_topic=global_config.llm_summary_by_topic,
)

90
src/plugins/memory_system/sample_distribution.py
View File

@@ -2,11 +2,12 @@ import numpy as np
from scipy import stats
from datetime import datetime, timedelta
class DistributionVisualizer:
def __init__(self, mean=0, std=1, skewness=0, sample_size=10):
"""
初始化分布可视化器
参数:
mean (float): 期望均值
std (float): 标准差
@@ -18,7 +19,7 @@ class DistributionVisualizer:
self.skewness = skewness
self.sample_size = sample_size
self.samples = None
def generate_samples(self):
"""生成具有指定参数的样本"""
if self.skewness == 0:
@@ -26,37 +27,28 @@ class DistributionVisualizer:
self.samples = np.random.normal(loc=self.mean, scale=self.std, size=self.sample_size)
else:
# 使用 scipy.stats 生成具有偏度的分布
self.samples = stats.skewnorm.rvs(a=self.skewness,
loc=self.mean,
scale=self.std,
size=self.sample_size)
self.samples = stats.skewnorm.rvs(a=self.skewness, loc=self.mean, scale=self.std, size=self.sample_size)
def get_weighted_samples(self):
"""获取加权后的样本数列"""
if self.samples is None:
self.generate_samples()
# 将样本值乘以样本大小
return self.samples * self.sample_size
def get_statistics(self):
"""获取分布的统计信息"""
if self.samples is None:
self.generate_samples()
return {
"均值": np.mean(self.samples),
"标准差": np.std(self.samples),
"实际偏度": stats.skew(self.samples)
}
return {"均值": np.mean(self.samples), "标准差": np.std(self.samples), "实际偏度": stats.skew(self.samples)}
class MemoryBuildScheduler:
def __init__(self,
n_hours1, std_hours1, weight1,
n_hours2, std_hours2, weight2,
total_samples=50):
def __init__(self, n_hours1, std_hours1, weight1, n_hours2, std_hours2, weight2, total_samples=50):
"""
初始化记忆构建调度器
参数:
n_hours1 (float): 第一个分布的均值（距离现在的小时数）
std_hours1 (float): 第一个分布的标准差（小时）
@@ -70,39 +62,31 @@ class MemoryBuildScheduler:
total_weight = weight1 + weight2
self.weight1 = weight1 / total_weight
self.weight2 = weight2 / total_weight
self.n_hours1 = n_hours1
self.std_hours1 = std_hours1
self.n_hours2 = n_hours2
self.std_hours2 = std_hours2
self.total_samples = total_samples
self.base_time = datetime.now()
def generate_time_samples(self):
"""生成混合分布的时间采样点"""
# 根据权重计算每个分布的样本数
samples1 = int(self.total_samples * self.weight1)
samples2 = self.total_samples - samples1
# 生成两个正态分布的小时偏移
hours_offset1 = np.random.normal(
loc=self.n_hours1,
scale=self.std_hours1,
size=samples1
)
hours_offset2 = np.random.normal(
loc=self.n_hours2,
scale=self.std_hours2,
size=samples2
)
hours_offset1 = np.random.normal(loc=self.n_hours1, scale=self.std_hours1, size=samples1)
hours_offset2 = np.random.normal(loc=self.n_hours2, scale=self.std_hours2, size=samples2)
# 合并两个分布的偏移
hours_offset = np.concatenate([hours_offset1, hours_offset2])
# 将偏移转换为实际时间戳（使用绝对值确保时间点在过去）
timestamps = [self.base_time - timedelta(hours=abs(offset)) for offset in hours_offset]
# 按时间排序（从最早到最近）
return sorted(timestamps)
@@ -111,54 +95,56 @@ class MemoryBuildScheduler:
timestamps = self.generate_time_samples()
return [int(t.timestamp()) for t in timestamps]
def print_time_samples(timestamps, show_distribution=True):
"""打印时间样本和分布信息"""
print(f"\n生成的{len(timestamps)}个时间点分布：")
print("序号".ljust(5), "时间戳".ljust(25), "距现在（小时）")
print("-" * 50)
now = datetime.now()
time_diffs = []
for i, timestamp in enumerate(timestamps, 1):
hours_diff = (now - timestamp).total_seconds() / 3600
time_diffs.append(hours_diff)
print(f"{str(i).ljust(5)} {timestamp.strftime('%Y-%m-%d %H:%M:%S').ljust(25)} {hours_diff:.2f}")
# 打印统计信息
print("\n统计信息：")
print(f"平均时间偏移：{np.mean(time_diffs):.2f}小时")
print(f"标准差：{np.std(time_diffs):.2f}小时")
print(f"最早时间：{min(timestamps).strftime('%Y-%m-%d %H:%M:%S')} ({max(time_diffs):.2f}小时前)")
print(f"最近时间：{max(timestamps).strftime('%Y-%m-%d %H:%M:%S')} ({min(time_diffs):.2f}小时前)")
if show_distribution:
# 计算时间分布的直方图
hist, bins = np.histogram(time_diffs, bins=40)
print("\n时间分布（每个*代表一个时间点）：")
for i in range(len(hist)):
if hist[i] > 0:
print(f"{bins[i]:6.1f}-{bins[i+1]:6.1f}小时: {'*' * int(hist[i])}")
print(f"{bins[i]:6.1f}-{bins[i + 1]:6.1f}小时: {'*' * int(hist[i])}")
# 使用示例
if __name__ == "__main__":
# 创建一个双峰分布的记忆调度器
scheduler = MemoryBuildScheduler(
n_hours1=12, # 第一个分布均值（12小时前）
std_hours1=8, # 第一个分布标准差
weight1=0.7, # 第一个分布权重 70%
n_hours2=36, # 第二个分布均值（36小时前）
std_hours2=24, # 第二个分布标准差
weight2=0.3, # 第二个分布权重 30%
total_samples=50 # 总共生成50个时间点
n_hours1=12, # 第一个分布均值（12小时前）
std_hours1=8, # 第一个分布标准差
weight1=0.7, # 第一个分布权重 70%
n_hours2=36, # 第二个分布均值（36小时前）
std_hours2=24, # 第二个分布标准差
weight2=0.3, # 第二个分布权重 30%
total_samples=50, # 总共生成50个时间点
)
# 生成时间分布
timestamps = scheduler.generate_time_samples()
# 打印结果，包含分布可视化
print_time_samples(timestamps, show_distribution=True)
# 打印时间戳数组
timestamp_array = scheduler.get_timestamp_array()
print("\n时间戳数组（Unix时间戳）：")
@@ -167,4 +153,4 @@ if __name__ == "__main__":
if i > 0:
print(", ", end="")
print(ts, end="")
print("]")
print("]")