配置

- id: dsh-memory
  name: '@restregular/dsh-memory'
  config:
    storeDir: '~/.dsh/memory-store'       # 建议绝对路径；默认 .dsh-memory
    scopes:
      global: true
      project: true
      session: true
      custom: { enabled: true, max: 200 } # 只统计自定义作用域
    retrieval:
      topK: 5
      recallN: 30
      vector:
        enabled: true
        endpoint: 'http://127.0.0.1:8078/v1/embeddings'
        model: bge-m3
        dims: 1024
        timeoutMs: 300                     # 交互检索预算，要紧
        writeTimeoutMs: 10000              # 保存时嵌整条记录
        backfillTimeoutMs: 30000
        backfillMaxChars: 8000
        spawn:
          enabled: true
          exe: '/path/to/llama-server.exe'
          model: '/path/to/bge-m3-Q4_K_M.gguf'
          cwd: '/path/to/llamacpp'         # 目录里要有 llama-server-impl.dll
          logFile: '/path/to/llama-server.log'
      llm: { enabled: false }
    limits: { maxRecordsPerScope: 2000, maxRecordChars: 4000 }
    secrets: { enabled: true, mode: refuse }   # refuse | redact
    tools: { defaultScope: session }            # session / project / global
    inject:
      enabled: true                             # 整个注入半边总开关
      extraScopes: ['handbook']                 # 可选：额外参与普通召回的 scope
      recall: { enabled: true, topK: 5, tokenBudget: 500 }
      impression: { enabled: true, chars: 120 }
      digest: { enabled: true, topK: 8, tokenBudget: 800, hintBudget: 160, hintMaxItems: 6 }
      policy: { enabled: true }
      rules: { enabled: true, maxItems: 10, tokenBudget: 300, delta: true }
      turnEnd: { enabled: false }
    panel: { route: '/memory/api' }

面板改过的值写入 /settings.json，优先级高于这里的配置。注入分层各自独立开关；inject.enabled: false 全关，但工具检索不受影响。

只有写了 endpoint、配了 spawn、或面板里装了本地后端，插件才会碰网络；否则一个请求都不发。

@restregular/dsh-memory

同名包的其他仓库

安装

说明文档

配置