Memento-Skills

Memento-Skills: Let Agents Design Agents

Deploy an agent with reusable, persistent, and extensible skills.

Memento-S Desktop App Is Live

The skill-centric agent runtime is now available as a desktop app.

You no longer need to start Memento-S from source. Windows and macOS users can install the app directly and try desktop chat, web research, multi-step task execution, skill calls, and reflection-based replanning.

During the launch period, new app users can receive a generous token package for long tasks, multi-turn collaboration, and agent reflection workflows. Check the app for the latest campaign details.

App Launch · What's New · Learning Results · Install · Quick Start · What Is This · Why It Matters · Ecosystem · Citation

English · 中文摘要

The accompanying technical study compares parameter training with deployment-time skill learning. Its research configuration keeps θ frozen and stores reusable behaviour in an external skill memory M.

Read–Execute–Reflect–Write architecture studied in the accompanying technical work. The public runtime supplies Qwen/keyword skill routing, execution, reflection and replanning, explicit skill creation, and guarded automatic skill repair.

What's New in v0.4.0

v0.4.0 prepares the public runtime for reproducible software review with consistent release metadata, explicit source licensing, installed-package checks, and clearer contributor and security guidance.

What's New in v0.3.8

v0.3.8 publishes the Memento-S desktop app as the recommended path for most users. Windows and macOS builds are now available directly from the README, with future product updates continuing through the app.

What's New in v0.3.0

v0.3.0 focuses on cleaner separation of concerns. The runtime is now organised into a dedicated infrastructure layer, a unified tool registry, and a long-lived agent profile system, so that core agent logic, infrastructure, and tools can evolve independently.

Infrastructure Layer (infra/)

A new top-level infra/ package isolates infrastructure concerns from core/ so that agent logic and platform code can evolve independently.

Unified Tool Registry (tools/)

Tooling has been promoted to a top-level tools/ package with a single ToolRegistry. The previous builtin/tools/ layer has been retired in favour of this unified surface.

Agent Profile System

core/agent_profile/ introduces persistent profiles that capture an agent's long-term identity, while a matching daemon/agent_profile/ package evolves those profiles in the background.

Dream Daemon

A new daemon/dream/ package runs background consolidation between sessions, similar to how reflection updates the skill library.

Shared Layer Expansion (shared/)

The shared/ package has grown beyond chat utilities into a broader foundation for cross-cutting helpers.

Runtime & Developer Experience

Refactors and Removals

The new layout supersedes several v0.2.0 modules. The v0.2.0 tool_bridge/ layer (runner, bridge, context, args_processor, result_processor) is replaced by the unified tools/ registry. builtin/tools/ (bash, file_ops, grep, python_repl, web) has been moved to tools/atomics/. The monolithic core/skill/execution/executor.py and core/context/manager.py, together with core/context/{block,scratchpad,runtime_state,memory}.py, have been split into tools/, infra/context/, and infra/memory/. IM startup no longer goes through middleware/im/gateway_starter.py and middleware/im/gateway/agent_worker.py; both are removed and replaced by a single EndpointService (server/endpoint/im/) that manages all IM channels and the agent worker. bootstrap.py and middleware/llm/llm_client.py have been reworked to align with this new wiring.

Compatibility Notes

• Modules previously imported from builtin.tools.* are now under tools.* (for example, tools.atomics, tools.mcp, tools.registry).
• Modules previously under core/shared/* (compact, memory, context wiring) now live under infra/*. Update any direct imports accordingly.
• core/manager/ has been removed; conversation and session management now lives in shared/chat/.
• IM channels are now started via EndpointService.start_channel(...) (server/endpoint/im/); direct use of gateway_starter / agent_worker no longer applies.

What's New in v0.2.0

v0.2.0 is a major architectural upgrade. The core agent, skill system, configuration layer, and deployment surfaces have all been redesigned or significantly extended compared to v0.1.0.

Core Architecture

Configuration System v2

Skill Ecosystem

IM Platform Integration (New)

A unified IM Gateway (middleware/im/gateway/) with AgentWorker, ConnectionManager, and platform-specific channels enables real-time message handling across all four platforms.

IM Gateway

GUI Enhancements

• Workspace browser — integrated file tree with drag-and-drop and in-place file operations.
• Session management — save, load, rename, and delete conversation history.
• Slash commands — /skills, /context, /compress, /feishu start|stop|status, and more.

Developer Experience

Operations

• howto.md — a standalone quick-start guide for local source-based deployment.
• requirements-dev.txt — separate dev dependencies.
• .github/ — CI/CD workflows.

Learning Results

The accompanying technical study evaluates a research configuration of Memento-Skills on two challenging benchmarks:

• HLE (Humanity's Last Exam) — a benchmark of extremely difficult questions spanning diverse academic disciplines, designed to probe the upper limits of frontier AI systems on expert-level reasoning and knowledge.

• GAIA (General AI Assistants) — a benchmark for evaluating general-purpose AI assistants on real-world tasks that require multi-step reasoning, web browsing, file handling, and tool use.

Results from the research configuration on two benchmarks. (a, b) depict task performance across reflective learning rounds on HLE and GAIA. (c, d) depict growth of the research skill memory and its semantic clusters.

The reported results improve over multiple learning rounds on HLE and GAIA, while the research skill library grows from a small set of atomic skills into a richer set of learned skills.

Core question. How can a deployable agent organise behaviour as reusable skills, reflect on execution, and persist newly authored capabilities? The accompanying technical study motivates the runtime's guarded automatic skill repair loop.

Reflect on execution

Step results inform whether the runtime continues, replans, or finalises.

Author reusable skills

The runtime can create missing skills and repair an attributed skill after a failed trajectory.

Run in the real world

Local execution, persistent state, CLI, GUI, and multi-platform IM integration make it deployable beyond a paper demo.

Key Features

What Is Memento-Skills?

Memento-Skills is a skill-centric agent runtime organised around skills as first-class units of capability. Skills are retrievable, executable, and persistent. Instead of treating tools as a flat pile of functions, Memento-Skills treats them as a library that can be routed, installed, authored, and reused over time.

The runtime combines step-boundary reflection and replanning with explicit skill authoring and guarded automatic write-back. Failed trajectories can repair the attributed skill without changing the underlying language model, while tests and directory snapshots prevent a bad candidate from replacing the active version.

What Makes It Different?

Memento-Skills exposes a Read -> Execute -> Reflect -> Write skill lifecycle.

This lifecycle keeps skill retrieval, execution, reflection, authoring, and persistence within one deployable runtime.

Memento-Skills vs OpenClaw

The two systems share a lot of DNA, but they are not centred on the same question.

• OpenClaw is more about getting an assistant to run in the real world.
• Memento-Skills is more about getting an agent to learn from the real world.

Shared Foundation

Key Differences

In one sentence: OpenClaw is about getting the assistant running; Memento-Skills is about getting the agent learning.

Deployment Surfaces

One Repo. One Learning Agent.

python -m venv .venv && source .venv/bin/activate && pip install -e . && memento doctor && memento agent

One-Click GUI Install

Download the pre-built desktop app — no Python or terminal needed. Install the package for your platform and start using Memento-S directly.

The desktop app is now the recommended path for most users. We will keep shipping product updates through the app, while this repository remains available for developers who want to inspect, extend, or run the source code.

Launch offer: new app users can receive a generous token package during the launch period. Open the app to check the latest campaign details.

For source-based development, continue with the developer quick start below.

Quick Start (Developer)

git clone https://github.com/Memento-Teams/Memento-Skills.git
cd Memento-Skills
python -m venv .venv
source .venv/bin/activate
pip install -e .

On first launch, ~/memento_s/config.json is created automatically. Fill in your model profile, then start the app:

memento doctor    # Check environment, dependencies, and config validity
memento agent     # Start an interactive agent session in the terminal
memento-gui       # Launch the desktop GUI with chat interface

Configuration system (v2)

Memento-Skills uses a three-layer configuration architecture (introduced in v0.2.0):

• System Config (system_config.json) — read-only defaults shipped with the codebase.
• User Config (~/memento_s/config.json) — persistent user customisation, read-write.
• Runtime Config — merged at startup (System + User), used at runtime.

When the config template updates across versions, the system automatically detects and merges new fields while preserving existing user values. Fields marked with x-managed-by: user are protected from auto-migration.

{
  "llm": {
    "active_profile": "default",
    "profiles": {
      "default": {
        "model": "openai/gpt-4o",
        "api_key": "your-api-key",
        "base_url": "https://api.openai.com/v1",
        "max_tokens": 8192,
        "temperature": 0.7,
        "timeout": 120
      }
    }
  },
  "env": {
    "TAVILY_API_KEY": "your-search-api-key"
  }
}

The model field uses the provider/model format, for example anthropic/claude-3.5-sonnet, openai/gpt-4o, or ollama/llama3. TAVILY_API_KEY is only required for web search.

The same profile system is also convenient for mainstream open-source model ecosystems, including Kimi / Moonshot, MiniMax, GLM / Zhipu, and other OpenAI-compatible endpoints.

Common commands

memento agent             # Interactive agent session
memento agent -m "..."    # Single-message mode
memento doctor            # Environment and config checks
memento verify            # Skill download / audit / execution validation
memento feishu            # Feishu IM bridge
memento wechat            # Optional WeChat integration (external SDK required)
memento im_status         # IM gateway status
memento-gui               # Desktop GUI

Supported LLM providers

Built-in Skills

The four Apache-compatible built-in skills are a starting point. Additional skills can be installed or authored after their licenses and provenance have been reviewed.

Verification and Contributions

GitHub Actions runs a deterministic smoke suite on Linux, macOS, and Windows, publishes coverage reports for that suite, and builds and inspects both the wheel and source archive. The artifact check prevents retired non-redistributable components from re-entering release packages.

See CONTRIBUTING.md for the development setup, exact local checks, integration-test policy, and third-party intake requirements.

Developer Notes

Project structure

Memento-Skills/
├── core/                  # Core agent framework
│   ├── memento_s/         # Agent orchestrator (4-stage ReAct + Finalize)
│   │   ├── phases/        # Intent, Planning, Execution/, Reflection, Finalize
│   │   └── skill_dispatch/ # Skill dispatch and tool routing
│   ├── skill/             # Skill framework
│   │   ├── loader/        # Skill discovery and loading
│   │   ├── retrieval/     # BM25 + vector hybrid retrieval
│   │   ├── execution/     # Sandbox execution + policies
│   │   ├── store/         # Skill persistence backends
│   │   ├── downloader/    # Skill download pipeline (new in v0.3.0)
│   │   ├── market.py      # Skill Market
│   │   └── gateway.py     # Unified skill gateway
│   ├── agent_profile/     # Agent profile system (new in v0.3.0)
│   ├── context/           # Bounded Context management
│   ├── protocol/          # Communication protocols
│   └── prompts/           # Prompt templates
├── infra/                 # Infrastructure layer (new in v0.3.0)
│   ├── memory/            # Long-term and session memory
│   ├── context/           # Context providers
│   ├── compact/           # Context compaction pipeline
│   ├── shared/            # Compact / extract helpers
│   └── service.py         # InfraService entry
├── tools/                 # Unified tool registry (new in v0.3.0)
│   ├── atomics/           # Atomic tools (bash, file, grep, web, repl, mcp)
│   ├── mcp/               # MCP client integration
│   └── registry.py        # ToolRegistry
├── middleware/             # Middleware layer
│   ├── config/            # Config v2 (three-layer isolation)
│   ├── llm/               # LLM client (litellm multi-provider)
│   ├── storage/           # SQLite + SQLAlchemy + vector storage
│   ├── im/                # IM platform middleware
│   ├── sandbox/           # uv sandbox
│   └── utils/             # Environment, path security
├── im/                    # IM platform integrations
│   ├── gateway/           # Gateway mode
│   ├── feishu/            # Feishu
│   ├── dingtalk/          # DingTalk
│   └── wecom/             # WeCom
├── gui/                   # Flet desktop GUI
├── cli/                   # Typer CLI
├── builtin/skills/        # Apache-compatible built-in skills (4)
├── shared/                # Shared layer (expanded in v0.3.0)
│   ├── chat/              # Session and conversation
│   ├── fs/, hooks/, schema/, security/, tools/  # Cross-cutting helpers
├── utils/                 # Shared helpers (runtime_requirements/ new in v0.3.0)
├── daemon/                # Background services
│   ├── agent_profile/     # Soul / user evolver (new in v0.3.0)
│   └── dream/             # Dream consolidation loop (new in v0.3.0)
├── tests/                 # Test suite
├── scripts/               # Build and deployment scripts
├── docs/                  # Documentation
├── bootstrap.py           # Application initialisation
├── version.py             # Version metadata
├── Figures/               # README figures
└── pyproject.toml         # Project configuration

Tech stack

FAQ

Memento Ecosystem

Memento-Skills is part of the broader Memento project family. Visit the links below to learn more about the full ecosystem and connect with the community.

Citation

If you find Memento-Skills useful in your research, please cite:

• Title: Memento-Skills: Let Agents Design Agents
• Authors: Huichi Zhou, Siyuan Guo, Anjie Liu, Zhongwei Yu, Ziqin Gong, Bowen Zhao, Zhixun Chen, Menglong Zhang, Yihang Chen, Jinsong Li, Runyu Yang, Qiangbin Liu, Xinlei Yu, Jianmin Zhou, Na Wang, Chunyang Sun, Jun Wang
• Paper: https://arxiv.org/abs/2603.18743

@article{zhou2026memento,
  title={Memento-Skills: Let Agents Design Agents},
  author={Zhou, Huichi and Guo, Siyuan and Liu, Anjie and Yu, Zhongwei and Gong, Ziqin and Zhao, Bowen and Chen, Zhixun and Zhang, Menglong and Chen, Yihang and Li, Jinsong and others},
  journal={arXiv preprint arXiv:2603.18743},
  year={2026}
}

Chinese Summary

点击展开中文摘要

Memento-Skills 把能力组织成可检索、可执行、可持久化的 skills。公开运行时围绕 Read -> Execute -> Reflect -> Write 生命周期提供 Qwen/关键词技能路由、执行、步骤反思与重新规划、显式技能创建，以及带自动归因、隔离更新、测试门禁和回滚的受控写回。

和 OpenClaw 相比，两者都具备 skills、工具调用、本地执行、持久化记忆和系统化部署能力，但关注点不同。OpenClaw 更偏向让 assistant 稳定接入真实世界；Memento-Skills 更强调统一的技能生命周期与持久化运行时。

v0.3.0 主要更新： 新增独立的 infra/ 基础设施层（memory / context / compact / service），将记忆、上下文、压缩等基础能力从 core/ 中拆出，便于独立演进；新增统一的 tools/ 工具注册中心（atomics、MCP、registry），取代原有 builtin/tools/；引入 Agent Profile 系统（core/agent_profile/ + daemon/agent_profile/），通过后台 evolver 持续维护 agent 的长期身份与用户画像；新增 daemon/dream/ 后台巩固循环，在会话间进行经验整理；扩展 shared/ 共享层（fs / hooks / schema / security / tools）；新增 utils/runtime_requirements/ 运行时依赖自动检测安装。

v0.2.0 主要更新： 核心架构采用 Bounded Context 重构，配置系统升级为三层隔离架构（System/User/Runtime），新增 Skill Market 支持云端技能市场，新增多平台 IM Gateway（飞书、钉钉、企业微信、微信），执行引擎拆分为细粒度子模块并增加 Tool Bridge、执行策略、错误恢复和循环检测机制，GUI 新增 Workspace 浏览器和会话管理，新增完整测试套件和构建脚本。

当前仓库已经具备比较完整的本地落地能力，包括 CLI、桌面 GUI、多平台 IM 桥接、本地 sandbox 和 skill 验证流程，因此它不仅适合 benchmark 式 agent，也适合继续往个人助理、长期运行和真实世界任务代理方向推进。

License

Memento-Skills is licensed under the Apache License 2.0. See THIRD_PARTY_NOTICES.md for bundled and optional third-party components.