config.yaml

openclaw openclaw官方 2026-04-09 1

OpenClaw 的持久记忆配置方法主要涉及 记忆存储、检索和管理 的设置,以下是关键配置步骤和方法：

config.yaml-第1张图片-OpenClaw开源下载|官方OpenClaw下载

配置记忆存储后端

OpenClaw 支持多种记忆存储方式：

A. 本地文件存储（JSON/文本）

  storage_type: "local"
  file_path: "./memory.json"
  # 或使用目录存储分片记忆
  directory: "./memory_chunks/"

B. 数据库存储

# 使用 PostgreSQL（带向量扩展）
memory:
  storage_type: "postgres"
  connection_string: "postgresql://user:pass@localhost:5432/memory_db"
  table_name: "agent_memories"
# 使用 MongoDB
memory:
  storage_type: "mongodb"
  connection_string: "mongodb://localhost:27017"
  database_name: "openclaw"
  collection_name: "memories"

C. 向量数据库（推荐）

# Pinecone
memory:
  storage_type: "pinecone"
  api_key: "${PINECONE_API_KEY}"
  environment: "us-west1-gcp"
  index_name: "openclaw-index"
# ChromaDB（本地）
memory:
  storage_type: "chromadb"
  path: "./chroma_data"
  collection_name: "agent_memory"

记忆检索策略配置

memory:
  retrieval:
    # 检索类型：similarity / time_weighted / 混合
    strategy: "hybrid"
    # 相似度检索设置
    similarity:
      top_k: 5
      threshold: 0.7  # 最小相似度
    # 时间加权检索
    recency_weight: 0.3
    importance_weight: 0.4
    relevance_weight: 0.3
    # 混合检索权重
    hybrid_weights:
      vector: 0.6
      keyword: 0.3
      time: 0.1

记忆处理流水线配置

memory:
  processing_pipeline:
    - name: "summarizer"
      model: "gpt-3.5-turbo"
      chunk_size: 2000  # 当记忆过长时自动总结
    - name: "embedder"
      model: "text-embedding-3-small"  # 或本地模型
    - name: "importance_scorer"
      # 自动计算记忆重要性（1-10）
    - name: "metadata_extractor"
      # 提取时间、实体、情感等元数据

记忆生命周期管理

memory:
  retention:
    max_items: 10000  # 最大记忆数量
    ttl_days: 30      # 过期时间
  compression:
    enabled: true
    # 当记忆超过阈值时自动压缩
    threshold: 1000
    compression_ratio: 0.3
  # 分层次存储
  hierarchical:
    levels:
      - name: "working_memory"
        capacity: 100
        ttl_hours: 24
      - name: "short_term"
        capacity: 1000
        ttl_days: 7
      - name: "long_term"
        capacity: 10000
        ttl_days: 365

完整配置示例

# openclaw_config.yaml
memory:
  enabled: true
  storage:
    type: "chromadb"
    persist_directory: "./memory_db"
    collection: "main_memory"
  embedding:
    model: "BAAI/bge-small-zh-v1.5"
    device: "cuda"  # 或 "cpu"
  retrieval:
    default_top_k: 7
    strategies:
      - type: "similarity"
        weight: 0.6
      - type: "recency"
        weight: 0.25
      - type: "importance"
        weight: 0.15
  summarization:
    model: "gpt-4-mini"
    trigger_length: 1500
    max_summary_length: 300
  indexing:
    auto_index: true
    batch_size: 100
    update_interval: 3600  # 每小时更新索引

环境变量配置

# .env 文件
MEMORY_STORAGE_TYPE=pinecone
PINECONE_API_KEY=your_key_here
OPENAI_API_KEY=sk-...  # 用于嵌入/
# 本地模型路径
LOCAL_EMBEDDING_MODEL_PATH=./models/bge-small

初始化代码示例

from openclaw import OpenClawAgent
from openclaw.memory import MemoryConfig
# 方法1：通过配置文件
agent = OpenClawAgent(
    config_file="config.yaml",
    memory_config="memory_config.yaml"
)
# 方法2：代码配置
memory_config = MemoryConfig(
    storage_type="chromadb",
    embedding_model="text-embedding-3-small",
    retrieval_strategy="hybrid",
    persist_directory="./memory_data"
)
agent = OpenClawAgent(
    name="assistant",
    memory_config=memory_config
)
# 使用记忆
agent.remember("用户喜欢咖啡")
memories = agent.recall("用户偏好")