XMemo

XMemo: Memory OS for AI Agents

XMemo is a user-owned Memory OS for AI agents: persistent memory, project context and session continuity with agent identity, provenance and governed access, shared across AI clients through MCP, the CLI and a public REST API surface.

English · 简体中文 | Documentation · Hosted MCP

Quick start · What this repo contains · Beyond MCP · Integrations · Connection modes · Plugins · Commands · Versioning · Security

What this repository contains

• [`@xmemo/client`](https://www.npmjs.com/package/@xmemo/client): the official xmemo CLI for setup, diagnostics, behavior profiles, and memory commands.
• MCP distribution: hosted Streamable HTTP configuration (https://xmemo.dev/mcp) and the local xmemo-mcp stdio server.
• XMemo Skill & native integrations: skills/xmemo for agent platforms and the native plugin index (src/plugins/index.json).
• Marketplace and registry metadata: canonical descriptors for MCP Registry (server.json), LobeHub (lhm.plugin.json), and AI agents (context7.json).

The hosted service implementation is private.

Beyond MCP

MCP is one access surface of XMemo, not the product boundary. XMemo provides a persistent, governed memory operating layer across AI tools and agents:

• [Persistent Memory](https://docs.xmemo.dev/docs/tools/remember): store, search, list, and soft-delete durable facts, preferences, and operational knowledge across sessions.
• [Context Recall](https://docs.xmemo.dev/docs/tools/recall-context): retrieve focused, relevant context on demand with token-budget controls and document expansions.
• [Session Continuity & Working State](https://docs.xmemo.dev/docs/guides/resume-and-handoff): restart snapshots and state save/restore so agents resume context across process restarts.
• [Project Context & Decisions](https://docs.xmemo.dev/docs/concepts/projects): track project-scoped facts, TODOs, and durable decisions across agent workflows.
• [Agent Identity & Provenance](https://docs.xmemo.dev/docs/concepts/provenance-attribution): track authorship and origin per memory record via agent instance attribution without conflating credentials.
• [Governed Access](https://docs.xmemo.dev/docs/concepts/scopes): scoped tokens, OAuth where the client supports it, soft/hard deletion and sensitive-memory handling (governance details).
• Native Integrations & Skill: dedicated integration for OpenClaw, Hermes, XMemo Skill (ClawHub), and DeepSeek DSH (available, released and pilot-tested; integration maturity Preview).
• Access Surfaces: unified via Model Context Protocol, the local CLI (xmemo), and the public REST API.

*Preview: Cloud Skills; Dream (off by default); Teams (Business plan not yet available). Knowledge Bases are available where enabled for the account.*

XMemo CLI

@xmemo/client is the official control plane for connecting AI tools to XMemo. It makes setup repeatable, keeps credentials out of project files, and gives every supported client a consistent path to durable, user-owned memory.

The package is deliberately small: the CLI runtime, safe client configuration, behavior profiles, XMemo skills, and marketplace metadata. Server code, databases, deployment files, logs, and internal operations remain outside the npm distribution.

Architecture

Why XMemo CLI

• One control plane — login, diagnostics, configuration, profiles, updates, and smoke checks share one predictable interface.
• Private by design — generated project configuration references a credential; it never embeds the credential value.
• Native where it matters — OpenClaw and Hermes use dedicated memory integrations instead of duplicating the same capability through MCP.
• Portable everywhere else — hosted Streamable HTTP MCP and local stdio cover modern editors, terminals, and agent runtimes.
• Safe automation — supported setup and removal paths offer preview, dry-run, or explicit confirmation before making changes.
• Small supply-chain surface — the npm package is governed by an explicit file allowlist and release provenance.

Quick start

Guided onboarding (xmemo init)

For an interactive first-run experience across account authentication, detected clients, agent behavior instructions, MCP server setup, skills, and plugins, run:

npm install -g @xmemo/client

xmemo init

Flags and options:

• xmemo init --dry-run: View the full onboarding plan without performing network calls or disk writes.
• xmemo init --yes: Automatically accept and apply all onboarding steps without interactive prompts.
• xmemo init --json: Emit structured JSON for the plan or result envelope.
• xmemo init --client ...: Restrict onboarding to specific clients (e.g., cursor, codex, claude-code).
• xmemo start: Alias for xmemo init with quick-start walkthrough steps.

Global installation exposes xmemo as the primary command, and also provides client and memory-os as aliases.

Manual step-by-step setup

xmemo account login

xmemo doctor xmemo setup codex xmemo status

Replace codex with your client. Preview a configuration before writing it:

xmemo setup cursor --dry-run

Running with npx

You can also run any CLI command directly without a global install via npx @xmemo/client :

# Check version or health

npx @xmemo/client --version npx @xmemo/client doctor

Guided onboarding without global install

npx @xmemo/client init

Install skill or run MCP stdio server

npx @xmemo/client skill install npx @xmemo/client mcp serve

[!TIP]

Start with xmemo init (or xmemo account login, xmemo doctor, and xmemo setup ).

Hand-edit MCP configuration only when a client has no verified setup path.

How the commands fit together

The XMemo CLI architecture is built on four core design principles:

1. Unified Resource Grammar (xmemo )

Every integration component is a first-class resource with predictable lifecycle actions:

• Composite commands: xmemo setup [ ...], xmemo uninstall [ ...], and xmemo status [ ] orchestrate these resources in a single step according to the client's declarative profile.
• Backward-compatible aliases: Familiar commands such as xmemo mcp add (alias for mcp install), xmemo profile uninstall (alias for profile remove), and xmemo skill uninstall (alias for skill remove) remain fully functional and print a helpful one-line hint in interactive terminals.

2. Unified Target Resolver

When no client is explicitly passed, the CLI uses a deterministic three-tier precedence resolution model:

1. Explicit flag or argument: --client , positional client argument, or --all.
2. Calling agent environment: Automatically identifies the calling agent runtime when running inside an agent session (e.g., CLAUDECODE / CLAUDE_CODE_ENTRYPOINT maps to claude-code, and CODEX_THREAD_ID / CODEX_SESSION_ID maps to codex).
3. Detected installed clients: Inspects local configuration paths and markers. If exactly one matching client is found, it is automatically selected; if multiple clients are found in interactive mode, an interactive picker is presented. The CLI never silently writes configuration to arbitrary unverified paths.

3. Plan, Confirm Once, Apply (PlanRunner)

Mutating commands follow a strict, atomic execution pattern:

1. Build Plan: Assemble an ordered sequence of actions across resources (e.g. plugin install followed by skill configuration).
2. Preview: Print the entire plan (including modified paths, commands, and unified diffs) to the terminal.
3. Confirm Once: Prompt [y/N] exactly once for the entire sequence. Re-running with --yes or -y bypasses the prompt; --dry-run displays the preview without mutation.
4. Apply Sequentially: Steps run in dependency order, stopping immediately upon first failure. Re-running an already configured client detects that all components are up to date and reports Nothing to do.

4. Declarative Client Registry

All client configurations, recipes, and capabilities are declared centrally in src/clients/registry.js. Command implementations are purely generic orchestrators with zero hardcoded client ID strings. Platforms can also be added dynamically at runtime via registerClient().

Supported integrations