OpenClaw 与 Hermes 的本地进化记忆系统Local evolving memory for OpenClaw and Hermes

快速开始Quick Start

安装器支持 macOS / Linux / Windows，会自动检查 Node.js 20+、下载 npm 包、重建原生依赖、写入 config.yaml，并按需启动宿主运行时。The installer supports macOS/Linux/Windows. It checks Node.js 20+, downloads the npm package, rebuilds native dependencies, writes config.yaml, and starts the host runtime when needed.

macOS / Linux：macOS / Linux:

# Latest from the repository
curl -fsSL https://raw.githubusercontent.com/MemTensor/MemOS/main/apps/memos-local-plugin/install.sh | bash

Windows (PowerShell)：Windows (PowerShell):

# Latest from the repository
irm https://raw.githubusercontent.com/MemTensor/MemOS/main/apps/memos-local-plugin/install.ps1 | iex

从源码目录安装或测试指定版本：Install from a source checkout or test a specific version:

cd apps/memos-local-plugin
bash install.sh --version 2.0.0-beta.11

# Or test a local package before publishing
npm pack
bash install.sh --version ./memtensor-memos-local-plugin-*.tgz

运行时路径Runtime Paths

运行时目录包含：Each runtime directory contains:

config.yaml        # only config file; chmod 600, secrets live here
data/memos.db      # SQLite: sessions, episodes, traces, policies, world, skills, feedback
skills/            # crystallized skill packages
logs/              # memos.log, error.log, audit.log, llm.jsonl, perf.jsonl, events.jsonl
daemon/            # bridge pid/port files, mostly for Hermes

分层架构Layered Architecture

Reflect2Evolve

新版算法把记忆拆成四个互相反馈的层级，并用 step-level 与 task-level 两类反馈持续更新。The new algorithm splits memory into four cooperating layers and continuously updates them from step-level and task-level feedback.

奖励回传Reward Backpropagation

V_T = R_human
V_t = alpha_t * R + (1 - alpha_t) * gamma * V_{t+1}
priority = value * time_decay

显式反馈、隐式结果和反思质量共同影响 trace 的价值，L2 候选池、L3 抽象和 Skill 生命周期都依赖这些信号。Explicit feedback, implicit outcomes, and reflection quality all affect trace value. L2 candidates, L3 abstractions, and Skill lifecycle decisions depend on these signals.

三层检索Three-Tier Retrieval

1. Tier 1 — Skill: 优先找到可直接调用的技能和 invocation guide。Find callable skills and invocation guides first.
2. Tier 2 — Trace + Episode: 检索具体证据、任务片段和历史操作。Retrieve concrete evidence, task slices, and historical actions.
3. Tier 3 — World Model: 补充项目级、环境级、长期约束级认知。Add project-level, environment-level, long-term cognition.

检索入口包括 turn start、tool-driven、skill invoke、sub-agent 和 repair。结果通过 RRF 融合、MMR 去冗余，再形成注入包。Entry points include turn start, tool-driven retrieval, skill invoke, sub-agent, and repair. Results are fused with RRF, diversified with MMR, and packed into an injection packet.

Viewer

配置Configuration

config.yaml 是唯一配置文件，敏感字段也写在这里，并由安装器设置为 chmod 600。也可以在 Viewer 的 Settings 页面编辑。config.yaml is the only config file. Secrets live there too, and the installer sets chmod 600. You can also edit it in Viewer Settings.

OpenClaw

version: 1
viewer:
  port: 18799
embedding:
  provider: local
  apiKey: ""
llm:
  provider: host
  apiKey: ""
  model: ""
hub:
  enabled: false
telemetry:
  enabled: true
logging:
  level: info
  detailedView: false

Hermes

version: 1
viewer:
  port: 18800
embedding:
  provider: local
  apiKey: ""
llm:
  provider: openai_compatible
  apiKey: ""   # required for Hermes unless you choose another real provider
  model: ""
hub:
  enabled: false
telemetry:
  enabled: true
logging:
  level: info
  detailedView: false

HTTP / Events

安全与本地优先Security and Local-First

• 数据库、技能包、日志和配置默认都在 ~/.<agent>/memos-plugin/。Database, skills, logs, and config live under ~/.<agent>/memos-plugin/.
• API key 不进入源码目录，不写入 URL，由 config loader 与 logger 做脱敏。API keys never enter the source directory or URLs; config loading and logging redact secrets.
• Viewer 默认绑定本机地址，每个宿主独立端口，避免跨宿主误写。Viewer binds locally by default; each host has an independent port to avoid cross-host writes.
• 遥测为匿名使用事件，可随时在配置中关闭。Telemetry is anonymous usage events and can be disabled in config.