MemOS Local Plugin | OpenClaw / Hermes 本地进化记忆

没有记忆的 Agent，每次都从零开始Without Memory, Every Task Starts from Zero

MemOS 为 OpenClaw 与 Hermes 提供同一套本地优先的分层记忆核心。MemOS gives both OpenClaw and Hermes the same local-first layered memory core.

💻

完全本地化Fully Local

记忆、技能、日志和配置都保存在本机运行时目录。Memory, skills, logs, and config stay in the local runtime directory.

🧠

全量可视化管理Full Visualization

Viewer 可查看 trace、policy、world model、skill、日志与配置。The Viewer exposes traces, policies, world models, skills, logs, and settings.

⚡

策略与技能进化Policy & Skill Evolution

高价值 trace 归纳成 L2 policy，成熟策略再结晶为可调用 Skill。High-value traces induce L2 policies, and mature policies crystallize into callable Skills.

💰

多模型后端Multiple Model Backends

Embedding 与 LLM 可分别选择本地、OpenAI 兼容、Gemini、Anthropic、Bedrock 等提供方。Embedding and LLM backends can be configured independently across local and cloud providers.

🤝

可选团队生态Optional Team Ecosystem

配置中保留 Hub 能力入口；默认关闭，不影响本地记忆主流程。Hub configuration is available as an optional surface and stays out of the local memory path by default.

🧬

宿主数据隔离Per-Host Isolation

OpenClaw 与 Hermes 使用独立运行时目录和 Viewer 端口，避免数据归属混淆。OpenClaw and Hermes use separate runtime homes and Viewer ports to keep ownership clear.

核心能力，驱动 Agent 持续进化Core Capabilities for Continuous Evolution

Reflect2Evolve 分层记忆Reflect2Evolve Layered Memory

每个回合先沉淀为 L1 trace，再由反馈信号驱动 L2 policy、L3 world model 和 Skill 生成。算法核心与宿主无关，因此 OpenClaw 和 Hermes 使用同一套演化逻辑。Turns become L1 traces first; feedback then drives L2 policies, L3 world models, and Skills. The algorithm core is host-agnostic, so OpenClaw and Hermes share the same evolution loop.

L1 TraceL2 PolicyL3 World ModelSkill反馈回传Reward Backprop

Task → Skill Evolution

L1 trace: action + observation + reflection + value
Reward: V_t = alpha_t * R + (1-alpha_t) * gamma * V_{t+1}

L2 policy: trigger + procedure + verification + boundary
L3 world: project/environment knowledge from L2 + L1
Skill: invocation guide + procedureJson

✓ Retrieval injects Skill → Trace/Episode → World Model

两个宿主，一个核心Two Hosts, One Core

OpenClaw 适配器在 TypeScript 进程内直接调用 core；Hermes 适配器通过 Python MemoryProvider 和 JSON-RPC bridge 访问同一个 core。适配器只处理宿主协议，算法、存储、检索和技能生命周期都留在共享核心里。The OpenClaw adapter calls core in-process from TypeScript; the Hermes adapter reaches the same core through a Python MemoryProvider and JSON-RPC bridge. Adapters handle host protocol only; storage, retrieval, and skill lifecycle stay shared.

OpenClaw TSHermes PythonJSON-RPC Bridge

Host Adapters

OpenClaw:
  before_prompt_build → core.onTurnStart()
  agent_end           → core.onTurnEnd()

Hermes:
  prefetch  → JSON-RPC turn.start
  sync_turn → JSON-RPC turn.end

✓ Same SQLite schema and retrieval pipeline

全量记忆可视化管理Full Memory Visualization

内置 Web 管理面板可查看 trace、policy、world model、skill、日志、配置和导入状态。OpenClaw 默认端口 18799，Hermes 默认端口 18800。The built-in dashboard shows traces, policies, world models, skills, logs, settings, and import state. OpenClaw uses port 18799; Hermes uses 18800.

127.0.0.1:18799

TracesPoliciesWorldSkillsLogsSettings

总记忆Total

1,284

今日Today

+47

任务Tasks

12

技能Skills

8

user帮我配置 Nginx 反向代理到 3000 端口Set up Nginx proxy to port 30002m

asst好的，创建 nginx 配置文件并写入 upstream 配置。Creating nginx config file and writing upstream block.2m

user还需要加 SSL 证书Also add SSL cert5m

从宿主事件到记忆注入的共享架构A Shared Architecture from Host Event to Memory Injection

OpenClaw 与 Hermes 的差异只存在于 adapter 层；core、server、viewer、SQLite schema 和算法事件保持一致。OpenClaw and Hermes differ only at the adapter layer; core, server, viewer, SQLite schema, and algorithm events stay shared.

架构层次Architecture Layers

1

Host AdapterHost Adapter

OpenClaw 使用进程内 TS 插件；Hermes 使用 Python provider + bridge。OpenClaw uses an in-process TS plugin; Hermes uses a Python provider plus bridge.

2

agent-contract

DTO、事件、错误码和 JSON-RPC 方法名在这里稳定下来。DTOs, events, errors, and JSON-RPC method names live here.

3

core

capture、reward、L1/L2/L3、skill、retrieval、storage、logger 都在共享核心里。Capture, reward, L1/L2/L3, skills, retrieval, storage, and logging live in the shared core.

4

server / viewer

HTTP + SSE 提供 Viewer 与调试入口，每个宿主有独立端口。HTTP + SSE power the Viewer and diagnostics, with one port per host.

60 秒上手Up and Running in 60 Seconds

macOS / Linux 安装器会自动检测 OpenClaw 与 Hermes，并写入对应运行时目录。The macOS/Linux installer auto-detects OpenClaw and Hermes and writes each runtime home.

1. 一键安装/升级1. Install

安装脚本会下载 npm 包、安装生产依赖、生成 config.yaml，并为 OpenClaw / Hermes 启动各自的 Viewer。
遇到安装问题？查看排查指南 →The script downloads the npm package, installs production dependencies, writes config.yaml, and starts the per-host Viewer.
Install issues? See troubleshooting guide →

# Step 1: 安装插件 & 启动（macOS/Linux）# Step 1: Install plugin & start (macOS/Linux)

$ curl -fsSL https://raw.githubusercontent.com/MemTensor/MemOS/main/apps/memos-local-plugin/install.sh | bash

2. 配置2. Config

网页面板：OpenClaw 默认 http://127.0.0.1:18799，Hermes 默认 http://127.0.0.1:18800。运行时配置写在各自的 config.yaml。Web panel: OpenClaw defaults to http://127.0.0.1:18799, Hermes to http://127.0.0.1:18800. Runtime config lives in each host's config.yaml.

127.0.0.1:18799

TracesPoliciesWorldSkillsLogsSettings

Embedding

Providerlocal Cloudoptional API Keyconfigured only for cloud providers

LLM

OpenClawhost Hermesopenai_compatible API Keyrequired for cloud providers

Runtime

OpenClaw~/.openclaw/memos-plugin Hermes~/.hermes/memos-plugin

Viewer Port18799

保存即生效Save to apply

version: 1
viewer:
  port: 18799  # OpenClaw; Hermes uses 18800
embedding:
  provider: local
  apiKey: ""
llm:
  provider: host  # Hermes: openai_compatible or another real provider
  apiKey: ""
  model: ""
hub:
  enabled: false
telemetry:
  enabled: true
logging:
  level: info

适配你的技术栈Works with Your Preferred Stack

Embedding 与 LLM 后端独立配置；无云端 key 时仍可使用本地 embedding 和宿主模型能力。Embedding and LLM backends are configured independently; local embedding and host LLM paths remain available where supported.

OpenAI

Anthropic

Gemini

Bedrock

Cohere

Voyage

Mistral

本地Local

宿主工具与Viewer 能力Host Tools and Viewer Capabilities

🔍

memos_search

三层检索Three-tier search

📄

memos_get

读取 trace / policy / world modelFetch trace / policy / world model

📜

memos_timeline

查看 episode 时间线Episode timeline

🌎

memos_environment

查询 L3 环境认知Query L3 world models

⚡

memos_skill_list

列出候选和活跃技能List candidate and active skills

📘

memos_skill_get

获取技能调用指南Fetch invocation guide

Team 团队共享中心Team Sharing Hub

多实例协作 —
让团队的 Agent 共同进化Multi-Instance Collaboration —
Your Team's Agents Evolve Together

OpenClaw 与 Hermes 默认各自本地隔离。显式开启 Team Sharing 后，多个实例可以通过 Hub 共享技能和可选 Trace 摘要；私有数据库、配置与日志仍留在本机。OpenClaw and Hermes stay isolated by default. When Team Sharing is explicitly enabled, multiple instances can share skills and optional trace excerpts through a Hub while private DBs, config, and logs remain local.

团队共享支持的协作方式How Team Sharing Fits the Local Core

1

默认隔离Isolated by Default

OpenClaw 与 Hermes 的数据库、技能包和日志默认互不共享。OpenClaw and Hermes keep DBs, skill bundles, and logs separate by default.

2

显式开启Explicit Opt-In

hub.enabled、hub.address 和 token 写入 config.yaml 后才加入团队。Instances join a team only after hub.enabled, hub.address, and tokens are configured.

3

技能优先共享Skill-First Sharing

适合共享已结晶 Skill 和可选 Trace 摘要，而不是直接合并私有数据库。Designed for sharing crystallized skills and optional trace excerpts, not merging private databases.

4

离线可用Local When Offline

Hub 不可用时，本地记忆、检索和 Skill 生命周期仍按本机流程运行。If the Hub is unavailable, local memory, retrieval, and skill lifecycle continue normally.

Import 原生记忆导入Native Memory Import

再续前缘 —
过往的记忆，不会丢失Reconnect —
Your Past Memories, Never Lost

Viewer 提供导入入口：OpenClaw 读取原生 session JSONL，Hermes 读取原生 MEMORY.md；也支持 MemOS JSON bundle 的导入导出。The Viewer exposes import paths for OpenClaw native session JSONL, Hermes native MEMORY.md, and MemOS JSON bundles.

🚀

按宿主导入Per-Host Import

OpenClaw 与 Hermes 只扫描当前 Viewer 所属宿主的数据源。OpenClaw and Hermes scan only the data source for the current Viewer host.

🧬

Bundle 往返Bundle Round-Trip

导出文件可重新导入，用于迁移设备或备份本地记忆。Exported bundles can be imported again for device migration or backup.

⏸️

批量处理Batched Processing

原生导入按 offset / limit 分批执行，避免一次性处理过多历史数据。Native import runs in offset / limit batches to avoid processing too much history at once.

⚡

保守写入Conservative Writes

导入入口写入 trace bundle；后续价值、策略和技能由核心算法按正常流程处理。Import writes trace bundles; value, policy, and skill evolution remain handled by the core pipeline.

让 OpenClaw 与 Hermes
越用越聪明 Give OpenClaw and Hermes
Lasting Intelligence

没有记忆的 Agent，每次都从零开始Without Memory, Every Task Starts from Zero

完全本地化Fully Local

全量可视化管理Full Visualization

策略与技能进化Policy & Skill Evolution

多模型后端Multiple Model Backends

可选团队生态Optional Team Ecosystem

宿主数据隔离Per-Host Isolation

核心能力，驱动 Agent 持续进化Core Capabilities for Continuous Evolution

Reflect2Evolve 分层记忆Reflect2Evolve Layered Memory

两个宿主，一个核心Two Hosts, One Core

全量记忆可视化管理Full Memory Visualization

从宿主事件到记忆注入的共享架构A Shared Architecture from Host Event to Memory Injection

架构层次Architecture Layers

Host AdapterHost Adapter

agent-contract

core

server / viewer

60 秒上手Up and Running in 60 Seconds

1. 一键安装/升级1. Install

2. 配置2. Config

适配你的技术栈Works with Your Preferred Stack

宿主工具与Viewer 能力Host Tools and Viewer Capabilities

memos_search

memos_get

memos_timeline

memos_environment

memos_skill_list

memos_skill_get

再续前缘 —
过往的记忆，不会丢失Reconnect —
Your Past Memories, Never Lost

按宿主导入Per-Host Import

Bundle 往返Bundle Round-Trip

批量处理Batched Processing

保守写入Conservative Writes

相关产品生态Related Product Ecosystem

让 OpenClaw 与 Hermes
共享进化记忆Give OpenClaw and Hermes
Shared Evolving Memory

让 OpenClaw 与 Hermes越用越聪明 Give OpenClaw and HermesLasting Intelligence

没有记忆的 Agent，每次都从零开始Without Memory, Every Task Starts from Zero

完全本地化Fully Local

全量可视化管理Full Visualization

策略与技能进化Policy & Skill Evolution

多模型后端Multiple Model Backends

可选团队生态Optional Team Ecosystem

宿主数据隔离Per-Host Isolation

核心能力，驱动 Agent 持续进化Core Capabilities for Continuous Evolution

Reflect2Evolve 分层记忆Reflect2Evolve Layered Memory

两个宿主，一个核心Two Hosts, One Core

全量记忆可视化管理Full Memory Visualization

从宿主事件到记忆注入的共享架构A Shared Architecture from Host Event to Memory Injection

架构层次Architecture Layers

Host AdapterHost Adapter

agent-contract

core

server / viewer

60 秒上手Up and Running in 60 Seconds

1. 一键安装/升级1. Install

2. 配置2. Config

适配你的技术栈Works with Your Preferred Stack

宿主工具与Viewer 能力Host Tools and Viewer Capabilities

memos_search

memos_get

memos_timeline

memos_environment

memos_skill_list

memos_skill_get

多实例协作 —让团队的 Agent 共同进化Multi-Instance Collaboration —Your Team's Agents Evolve Together

团队共享支持的协作方式How Team Sharing Fits the Local Core

默认隔离Isolated by Default

显式开启Explicit Opt-In

技能优先共享Skill-First Sharing

离线可用Local When Offline

再续前缘 —过往的记忆，不会丢失Reconnect —Your Past Memories, Never Lost

按宿主导入Per-Host Import

Bundle 往返Bundle Round-Trip

批量处理Batched Processing

保守写入Conservative Writes

相关产品生态Related Product Ecosystem

让 OpenClaw 与 Hermes共享进化记忆Give OpenClaw and HermesShared Evolving Memory

让 OpenClaw 与 Hermes
越用越聪明 Give OpenClaw and Hermes
Lasting Intelligence

多实例协作 —
让团队的 Agent 共同进化Multi-Instance Collaboration —
Your Team's Agents Evolve Together

再续前缘 —
过往的记忆，不会丢失Reconnect —
Your Past Memories, Never Lost

让 OpenClaw 与 Hermes
共享进化记忆Give OpenClaw and Hermes
Shared Evolving Memory