diff --git a/src/memory_graph/config.py b/src/memory_graph/config.py index 054200807..a765f3fc1 100644 --- a/src/memory_graph/config.py +++ b/src/memory_graph/config.py @@ -66,6 +66,41 @@ class StorageConfig: class MemoryGraphConfig: """记忆图系统总配置""" + # 基础配置 + enable: bool = True # 是否启用记忆图系统 + data_dir: Path = field(default_factory=lambda: Path("data/memory_graph")) + + # 向量存储配置 + vector_collection_name: str = "memory_nodes" + vector_db_path: Path = field(default_factory=lambda: Path("data/memory_graph/chroma_db")) + + # 检索配置 + search_top_k: int = 10 + search_min_importance: float = 0.3 + search_similarity_threshold: float = 0.5 + enable_query_optimization: bool = True + + # 整合配置 + consolidation_enabled: bool = True + consolidation_interval_hours: float = 1.0 + consolidation_similarity_threshold: float = 0.85 + consolidation_time_window_hours: int = 24 + + # 遗忘配置 + forgetting_enabled: bool = True + forgetting_activation_threshold: float = 0.1 + forgetting_min_importance: float = 0.8 + + # 激活配置 + activation_decay_rate: float = 0.9 + activation_propagation_strength: float = 0.5 + activation_propagation_depth: int = 1 + + # 性能配置 + max_memory_nodes_per_memory: int = 10 + max_related_memories: int = 5 + + # 旧配置（向后兼容） consolidation: ConsolidationConfig = field(default_factory=ConsolidationConfig) retrieval: RetrievalConfig = field(default_factory=RetrievalConfig) node_merger: NodeMergerConfig = field(default_factory=NodeMergerConfig) @@ -89,10 +124,74 @@ class MemoryGraphConfig: enable_debug_logging: bool = False enable_visualization: bool = False # 是否启用记忆可视化 + @classmethod + def from_bot_config(cls, bot_config) -> MemoryGraphConfig: + """从bot_config加载配置""" + try: + # 尝试获取新配置 + if hasattr(bot_config, 'memory_graph'): + mg_config = bot_config.memory_graph + + config = cls( + enable=getattr(mg_config, 'enable', True), + data_dir=Path(getattr(mg_config, 'data_dir', 'data/memory_graph')), + vector_collection_name=getattr(mg_config, 'vector_collection_name', 'memory_nodes'), + vector_db_path=Path(getattr(mg_config, 'vector_db_path', 'data/memory_graph/chroma_db')), + search_top_k=getattr(mg_config, 'search_top_k', 10), + search_min_importance=getattr(mg_config, 'search_min_importance', 0.3), + search_similarity_threshold=getattr(mg_config, 'search_similarity_threshold', 0.5), + enable_query_optimization=getattr(mg_config, 'enable_query_optimization', True), + consolidation_enabled=getattr(mg_config, 'consolidation_enabled', True), + consolidation_interval_hours=getattr(mg_config, 'consolidation_interval_hours', 1.0), + consolidation_similarity_threshold=getattr(mg_config, 'consolidation_similarity_threshold', 0.85), + consolidation_time_window_hours=getattr(mg_config, 'consolidation_time_window_hours', 24), + forgetting_enabled=getattr(mg_config, 'forgetting_enabled', True), + forgetting_activation_threshold=getattr(mg_config, 'forgetting_activation_threshold', 0.1), + forgetting_min_importance=getattr(mg_config, 'forgetting_min_importance', 0.8), + activation_decay_rate=getattr(mg_config, 'activation_decay_rate', 0.9), + activation_propagation_strength=getattr(mg_config, 'activation_propagation_strength', 0.5), + activation_propagation_depth=getattr(mg_config, 'activation_propagation_depth', 1), + max_memory_nodes_per_memory=getattr(mg_config, 'max_memory_nodes_per_memory', 10), + max_related_memories=getattr(mg_config, 'max_related_memories', 5), + ) + + return config + else: + # 没有找到memory_graph配置，使用默认值 + return cls() + + except Exception as e: + import logging + logger = logging.getLogger(__name__) + logger.warning(f"从bot_config加载memory_graph配置失败，使用默认配置: {e}") + return cls() + @classmethod def from_dict(cls, config_dict: Dict) -> MemoryGraphConfig: """从字典创建配置""" return cls( + # 新配置字段 + enable=config_dict.get("enable", True), + data_dir=Path(config_dict.get("data_dir", "data/memory_graph")), + vector_collection_name=config_dict.get("vector_collection_name", "memory_nodes"), + vector_db_path=Path(config_dict.get("vector_db_path", "data/memory_graph/chroma_db")), + search_top_k=config_dict.get("search_top_k", 10), + search_min_importance=config_dict.get("search_min_importance", 0.3), + search_similarity_threshold=config_dict.get("search_similarity_threshold", 0.5), + enable_query_optimization=config_dict.get("enable_query_optimization", True), + consolidation_enabled=config_dict.get("consolidation_enabled", True), + consolidation_interval_hours=config_dict.get("consolidation_interval_hours", 1.0), + consolidation_similarity_threshold=config_dict.get("consolidation_similarity_threshold", 0.85), + consolidation_time_window_hours=config_dict.get("consolidation_time_window_hours", 24), + forgetting_enabled=config_dict.get("forgetting_enabled", True), + forgetting_activation_threshold=config_dict.get("forgetting_activation_threshold", 0.1), + forgetting_min_importance=config_dict.get("forgetting_min_importance", 0.8), + activation_decay_rate=config_dict.get("activation_decay_rate", 0.9), + activation_propagation_strength=config_dict.get("activation_propagation_strength", 0.5), + activation_propagation_depth=config_dict.get("activation_propagation_depth", 1), + max_memory_nodes_per_memory=config_dict.get("max_memory_nodes_per_memory", 10), + max_related_memories=config_dict.get("max_related_memories", 5), + # 旧配置字段（向后兼容） consolidation=ConsolidationConfig(**config_dict.get("consolidation", {})), retrieval=RetrievalConfig(**config_dict.get("retrieval", {})), node_merger=NodeMergerConfig(**config_dict.get("node_merger", {})), diff --git a/src/memory_graph/manager.py b/src/memory_graph/manager.py index 3c54e46b9..45cf32c4a 100644 --- a/src/memory_graph/manager.py +++ b/src/memory_graph/manager.py @@ -68,10 +68,10 @@ class MemoryManager: self._initialized = False self._last_maintenance = datetime.now() self._maintenance_task: Optional[asyncio.Task] = None - self._maintenance_interval_hours = 1 # 默认每小时执行一次维护 + self._maintenance_interval_hours = self.config.consolidation_interval_hours # 从配置读取 self._maintenance_schedule_id: Optional[str] = None # 调度任务ID - logger.info(f"记忆管理器已创建 (data_dir={data_dir})") + logger.info(f"记忆管理器已创建 (data_dir={data_dir}, enable={self.config.enable})") async def initialize(self) -> None: """ @@ -556,7 +556,7 @@ class MemoryManager: # 计算时间衰减 last_access_dt = datetime.fromisoformat(last_access) hours_passed = (now - last_access_dt).total_seconds() / 3600 - decay_factor = 0.9 ** (hours_passed / 24) # 每天衰减 10% + decay_factor = self.config.activation_decay_rate ** (hours_passed / 24) current_activation = activation_info.get("level", 0.0) * decay_factor else: current_activation = 0.0 @@ -573,12 +573,15 @@ class MemoryManager: memory.metadata["activation"] = activation_info memory.last_accessed = now - # 激活传播：激活相关记忆（强度减半） + # 激活传播：激活相关记忆 if strength > 0.1: # 只有足够强的激活才传播 - related_memories = self._get_related_memories(memory_id) - propagation_strength = strength * 0.5 + related_memories = self._get_related_memories( + memory_id, + max_depth=self.config.activation_propagation_depth + ) + propagation_strength = strength * self.config.activation_propagation_strength - for related_id in related_memories[:5]: # 最多传播到 5 个相关记忆 + for related_id in related_memories[:self.config.max_related_memories]: await self.activate_memory(related_id, propagation_strength) # 保存更新 @@ -677,7 +680,7 @@ class MemoryManager: continue # 跳过高重要性记忆 - if memory.importance >= 0.8: + if memory.importance >= self.config.forgetting_min_importance: continue # 计算当前激活度 @@ -875,15 +878,19 @@ class MemoryManager: } # 1. 记忆整理（合并相似记忆） - consolidate_result = await self.consolidate_memories( - similarity_threshold=0.85, - time_window_hours=24 - ) - result["consolidated"] = consolidate_result.get("merged_count", 0) + if self.config.consolidation_enabled: + consolidate_result = await self.consolidate_memories( + similarity_threshold=self.config.consolidation_similarity_threshold, + time_window_hours=self.config.consolidation_time_window_hours + ) + result["consolidated"] = consolidate_result.get("merged_count", 0) # 2. 自动遗忘 - forgotten_count = await self.auto_forget_memories(threshold=0.1) - result["forgotten"] = forgotten_count + if self.config.forgetting_enabled: + forgotten_count = await self.auto_forget_memories( + threshold=self.config.forgetting_activation_threshold + ) + result["forgotten"] = forgotten_count # 3. 清理非常旧的已遗忘记忆（可选） # TODO: 实现清理逻辑 diff --git a/src/memory_graph/manager_singleton.py b/src/memory_graph/manager_singleton.py index 0f2fd3b5c..55e1c026f 100644 --- a/src/memory_graph/manager_singleton.py +++ b/src/memory_graph/manager_singleton.py @@ -19,12 +19,16 @@ _memory_manager: Optional[MemoryManager] = None _initialized: bool = False -async def initialize_memory_manager(data_dir: Optional[Path | str] = None) -> MemoryManager: +async def initialize_memory_manager( + data_dir: Optional[Path | str] = None, + config = None, +) -> Optional[MemoryManager]: """ 初始化全局 MemoryManager Args: data_dir: 数据目录，默认使用 data/memory_graph + config: MemoryGraphConfig 或 bot_config 实例 Returns: MemoryManager 实例 @@ -36,14 +40,40 @@ async def initialize_memory_manager(data_dir: Optional[Path | str] = None) -> Me return _memory_manager try: + from src.memory_graph.config import MemoryGraphConfig + + # 处理配置 + if config is None: + # 尝试从全局配置加载 + try: + from src.config.config import global_config + memory_config = MemoryGraphConfig.from_bot_config(global_config) + logger.info("从 bot_config 加载 memory_graph 配置") + except Exception as e: + logger.warning(f"无法从 bot_config 加载配置，使用默认配置: {e}") + memory_config = MemoryGraphConfig() + elif isinstance(config, MemoryGraphConfig): + memory_config = config + else: + # 假设是 bot_config + memory_config = MemoryGraphConfig.from_bot_config(config) + + # 检查是否启用 + if not memory_config.enable: + logger.info("记忆图系统已在配置中禁用") + _initialized = False + _memory_manager = None + return None + + # 处理数据目录 if data_dir is None: - data_dir = Path("data/memory_graph") + data_dir = memory_config.data_dir elif isinstance(data_dir, str): data_dir = Path(data_dir) logger.info(f"正在初始化全局 MemoryManager (data_dir={data_dir})...") - _memory_manager = MemoryManager(data_dir=data_dir) + _memory_manager = MemoryManager(config=memory_config, data_dir=data_dir) await _memory_manager.initialize() _initialized = True diff --git a/template/bot_config_template.toml b/template/bot_config_template.toml index dc0123b98..c3414460f 100644 --- a/template/bot_config_template.toml +++ b/template/bot_config_template.toml @@ -235,95 +235,59 @@ enable_emotion_analysis = false # 是否启用表情包感情关键词二次识 emoji_selection_mode = "emotion" max_context_emojis = 30 # 每次随机传递给LLM的表情包详细描述的最大数量，0为全部 -[memory] -enable_memory = true # 是否启用记忆系统 -memory_build_interval = 600 # 记忆构建间隔（秒）。间隔越低，学习越频繁，但可能产生更多冗余信息 +# ==================== 记忆图系统配置 (Memory Graph System) ==================== +# 新一代记忆系统：基于知识图谱 + 语义向量的混合记忆架构 +# 替代旧的 enhanced memory 系统 -# === 记忆采样系统配置 === -memory_sampling_mode = "immediate" # 记忆采样模式：'immediate'(即时采样), 'hippocampus'(海马体定时采样) or 'all'(双模式) +[memory_graph] +# === 基础配置 === +enable = true # 是否启用记忆图系统 +data_dir = "data/memory_graph" # 记忆数据存储目录 -# 海马体双峰采样配置 -enable_hippocampus_sampling = true # 启用海马体双峰采样策略 -hippocampus_sample_interval = 1800 # 海马体采样间隔（秒，默认30分钟） -hippocampus_sample_size = 30 # 海马体采样样本数量 -hippocampus_batch_size = 10 # 海马体批量处理大小 -hippocampus_distribution_config = [12.0, 8.0, 0.7, 48.0, 24.0, 0.3] # 海马体双峰分布配置：[近期均值(h), 近期标准差(h), 近期权重, 远期均值(h), 远期标准差(h), 远期权重] +# === 向量存储配置 === +vector_collection_name = "memory_nodes" # 向量集合名称 +vector_db_path = "data/memory_graph/chroma_db" # 向量数据库路径 (使用独立的chromadb实例) -# 即时采样配置 -precision_memory_reply_threshold = 0.5 # 精准记忆回复阈值（0-1），高于此值的对话将立即构建记忆 +# === 记忆检索配置 === +search_top_k = 10 # 默认检索返回数量 +search_min_importance = 0.3 # 最小重要性阈值 (0.0-1.0) +search_similarity_threshold = 0.5 # 向量相似度阈值 -min_memory_length = 10 # 最小记忆长度 -max_memory_length = 500 # 最大记忆长度 -memory_value_threshold = 0.5 # 记忆价值阈值，低于该值的记忆会被丢弃 -vector_similarity_threshold = 0.7 # 向量相似度阈值 -semantic_similarity_threshold = 0.6 # 语义重排阶段的最低匹配阈值 +# 智能查询优化 +enable_query_optimization = true # 启用查询优化（使用小模型优化搜索查询） -metadata_filter_limit = 100 # 元数据过滤阶段返回数量上限 -vector_search_limit = 50 # 向量搜索阶段返回数量上限 -semantic_rerank_limit = 20 # 语义重排阶段返回数量上限 -final_result_limit = 10 # 综合筛选后的最终返回数量 +# === 记忆整合配置 === +consolidation_enabled = true # 是否启用记忆整合 +consolidation_interval_hours = 1.0 # 整合任务执行间隔（小时） +consolidation_similarity_threshold = 0.85 # 相似记忆合并阈值 +consolidation_time_window_hours = 24 # 整合时间窗口（小时） -vector_weight = 0.4 # 综合评分中向量相似度的权重 -semantic_weight = 0.3 # 综合评分中语义匹配的权重 -context_weight = 0.2 # 综合评分中上下文关联的权重 -recency_weight = 0.1 # 综合评分中时效性的权重 +# === 记忆遗忘配置 === +forgetting_enabled = true # 是否启用自动遗忘 +forgetting_activation_threshold = 0.1 # 激活度阈值（低于此值的记忆会被遗忘） +forgetting_min_importance = 0.8 # 最小保护重要性（高于此值的记忆不会被遗忘） -fusion_similarity_threshold = 0.85 # 记忆融合时的相似度阈值 -deduplication_window_hours = 24 # 记忆去重窗口（小时） +# === 记忆激活配置 === +activation_decay_rate = 0.9 # 激活度衰减率（每天衰减10%） +activation_propagation_strength = 0.5 # 激活传播强度（传播到相关记忆的激活度比例） +activation_propagation_depth = 1 # 激活传播深度（最多传播几层） -# 智能遗忘机制配置 (新增) -enable_memory_forgetting = true # 是否启用智能遗忘机制 -forgetting_check_interval_hours = 24 # 遗忘检查间隔（小时） +# === 性能配置 === +max_memory_nodes_per_memory = 10 # 每条记忆最多包含的节点数 +max_related_memories = 5 # 激活传播时最多影响的相关记忆数 -# 遗忘阈值配置 -base_forgetting_days = 30.0 # 基础遗忘天数 -min_forgetting_days = 7.0 # 最小遗忘天数（重要记忆也会被保留的最少天数） -max_forgetting_days = 365.0 # 最大遗忘天数（普通记忆最长保留天数） +# ==================== 旧记忆系统配置 (已弃用) ==================== +# 注意：以下配置仅用于向后兼容，新系统不使用这些配置 +# 旧的 enhanced memory 系统已被 memory_graph 系统取代 -# 重要程度权重 - 不同重要程度的额外保护天数 -critical_importance_bonus = 45.0 # 关键重要性额外天数 -high_importance_bonus = 30.0 # 高重要性额外天数 -normal_importance_bonus = 15.0 # 一般重要性额外天数 -low_importance_bonus = 0.0 # 低重要性额外天数 +[memory_legacy] +# 旧系统已禁用，所有配置保留仅供参考 +enable_legacy_memory = false # 旧记忆系统已禁用 -# 置信度权重 - 不同置信度的额外保护天数 -verified_confidence_bonus = 30.0 # 已验证置信度额外天数 -high_confidence_bonus = 20.0 # 高置信度额外天数 -medium_confidence_bonus = 10.0 # 中等置信度额外天数 -low_confidence_bonus = 0.0 # 低置信度额外天数 - -# 激活频率权重 -activation_frequency_weight = 0.5 # 每次激活增加的天数权重 -max_frequency_bonus = 10.0 # 最大激活频率奖励天数 - -# 休眠机制 -dormant_threshold_days = 90 # 休眠状态判定天数（超过此天数未访问的记忆进入休眠状态） - -# Vector DB存储配置 (新增 - 替代JSON存储) -enable_vector_memory_storage = true # 启用Vector DB存储 -enable_llm_instant_memory = true # 启用基于LLM的瞬时记忆 -enable_vector_instant_memory = true # 启用基于向量的瞬时记忆 -instant_memory_max_collections = 100 # 瞬时记忆最大集合数 -instant_memory_retention_hours = 0 # 瞬时记忆保留时间（小时），0表示不基于时间清理 - -# Vector DB配置 -vector_db_similarity_threshold = 0.5 # Vector DB相似度阈值 (推荐范围: 0.5-0.6, 过高会导致检索不到结果) -vector_db_search_limit = 20 # Vector DB单次搜索返回的最大结果数 -vector_db_batch_size = 100 # 批处理大小 (批量存储记忆时每批处理的记忆条数) -vector_db_enable_caching = true # 启用内存缓存 -vector_db_cache_size_limit = 1000 # 缓存大小限制 (内存缓存最多保存的记忆条数) -vector_db_auto_cleanup_interval = 3600 # 自动清理间隔（秒） -vector_db_retention_hours = 720 # 记忆保留时间（小时，默认30天） - -# 多阶段召回配置（可选） -# 取消注释以启用更严格的粗筛，适用于大规模记忆库（>10万条） -# memory_importance_threshold = 0.3 # 重要性阈值（过滤低价值记忆，范围0.0-1.0） -# memory_recency_days = 30 # 时间范围（只搜索最近N天的记忆，0表示不限制） - -# Vector DB配置 (ChromaDB) +# Vector DB配置 (ChromaDB) - 保留用于其他系统 [vector_db] type = "chromadb" # Vector DB类型 -path = "data/chroma_db" # Vector DB数据路径 +path = "data/chroma_db" # Vector DB数据路径（用于其他系统，非memory_graph） [vector_db.settings] anonymized_telemetry = false # 禁用匿名遥测