Self-learning AI Agent Framework with 6-Layer Evolving Memory
Build AI employees that get smarter every day — not chatbots that forget everything.
Homepage · Quick Start · Architecture · 中国版 · DeepBrain 独立引擎
Leaper Agent = Hermes Agent (battle-tested open-source agent runtime) + DeepBrain 6-Layer Memory Evolution Engine. We track Hermes upstream releases for maximum international compatibility, and layer our self-evolving memory system on top — giving agents the ability to continuously learn, distill, and govern knowledge across conversations.
Leaper Agent = Hermes Agent（成熟开源 Agent 运行时）+ DeepBrain 六层记忆进化引擎。基座同步 Hermes 版本更新，保证海外兼容性最大化；差异化核心是我们的自进化记忆系统 —— 让 AI 在持续交互中自动学习、提炼、治理知识。不是聊天机器人，是会成长的 AI 员工。
| Capability | LangChain | CrewAI | AutoGPT | Mem0 | Leaper Agent |
|---|---|---|---|---|---|
| Memory persistence | ❌ DIY | ❌ None | Short-term | ✅ Key-value | ✅ 6-layer evolution |
| Auto knowledge extraction | ❌ | ❌ | ❌ | ✅ Basic | ✅ 4Gate filtering |
| Knowledge governance (conflict/decay/merge) | ❌ | ❌ | ❌ | ❌ | ✅ L3 governance |
| User profiling | ❌ | ❌ | ❌ | Partial | ✅ L4 multi-dim |
| Consistency guard (regression/anomaly) | ❌ | ❌ | ❌ | ❌ | ✅ L5 guardian |
| Multi-agent isolation | ❌ | Partial | ❌ | ❌ | ✅ 1-process N-agents |
| Ready-to-use templates | ❌ | ❌ | ❌ | ❌ | ✅ 10 CXO + 140 industry |
| Multi-platform (Telegram/Discord/飞书/钉钉) | ❌ | ❌ | CLI | ❌ | ✅ 6 channels |
TL;DR: Other frameworks help you call LLMs. Leaper helps you build AI that remembers, learns, and grows.
pip install leaper-agent
leaper init # Configure API key & proxy
leaper create ceo-coach --bot-token YOUR_TOKEN # Create agent from template
leaper start ceo-coach # LaunchThat's it. Open Telegram, talk to your bot — it learns from day one.
The core engine that makes Leaper agents get smarter over time.
┌─────────────────────────────────────────────────────────────────┐
│ L0 Raw Storage 对话 / 文档 / 输入的原始数据 │
├─────────────────────────────────────────────────────────────────┤
│ L1 Structured Extract 4Gate 门控 → 只留有价值的信息 │
│ ┌──────────┬──────────┬──────────┬──────────┐ │
│ │ Novelty │Actionable│ Durable │ Relevant │ │
│ └──────────┴──────────┴──────────┴──────────┘ │
├─────────────────────────────────────────────────────────────────┤
│ L2 Skill Synthesis 跨对话聚类 · 知识合并 · 去重 │
├─────────────────────────────────────────────────────────────────┤
│ L3 Knowledge Gov. 冲突检测 · 淘汰 · 晋升 · 漂移检测 │
├─────────────────────────────────────────────────────────────────┤
│ L4 User Profiling 多维特征建模 · 偏好学习 │
├─────────────────────────────────────────────────────────────────┤
│ L5 Consistency Guard 回归测试 · 时间衰减 · 异常检测 │
└─────────────────────────────────────────────────────────────────┘
Not everything is worth remembering. Every piece of information passes through 4 gates:
| Gate | Function | Example |
|---|---|---|
| Novelty | Skip if already known | "Nice weather" → ❌ |
| Actionable | Skip if no value | "Hmm ok" → ❌ |
| Durable | TTL-classify ephemeral info | "Meeting tomorrow" → short-term |
| Relevant | Skip if off-topic | Idle chat noise → ❌ |
┌───────────────────────────────────────────────────────┐
│ Leaper Gateway │
│ Route · Auth · Session Mgmt │
├───────┬───────┬───────┬───────┬───────┬───────────────┤
│ CS │ Sales │ Ops │ Coach │ ... │ Custom │
│ Agent │ Agent │ Agent │ Agent │ │ Agents │
├───────┴───────┴───────┴───────┴───────┴───────────────┤
│ DeepBrain 6-Layer Engine │
│ (Memory isolated per agent, shared DB) │
├───────────────────────────────────────────────────────┤
│ Tool System: File · Terminal · Search · Browser · │
│ Code Exec · MCP Protocol │
├───────────────────────────────────────────────────────┤
│ LLM: OpenAI · Claude · DeepSeek · Qwen · Zhipu · │
│ Moonshot · Ollama │
├───────────────────────────────────────────────────────┤
│ Channels: Telegram · Discord · 飞书 · 钉钉 · │
│ 企业微信 · REST API │
└───────────────────────────────────────────────────────┘
Key design decisions:
- 1 process, N agents — shared infra, isolated memory, low overhead
- MCP protocol — extensible tool ecosystem
- Multi-LLM — switch providers without code changes
- Usage billing + cost optimization built in
10 CXO role templates + 140 industry-specific variants:
leaper templates list # Browse all
leaper create my-cfo --template cfo --bot-token TOKEN # CFO advisor
leaper create my-cs --template customer-success --industry ecommerce --bot-token TOKEN
leaper start --all # Launch all| Role | What It Does |
|---|---|
| CEO Coach | Strategic thinking, decision frameworks, OKR review |
| CTO | Tech stack review, architecture decisions, trend analysis |
| CFO | Financial analysis, cost control, budget planning |
| CMO | Marketing strategy, user growth, competitor analysis |
| Customer Success | Client history, renewal alerts, churn prediction |
| ... | 140+ industry variants (e-commerce, SaaS, healthcare...) |
Enterprise: Onboarding bots that learn company policies · CS agents that remember every client · Tech advisors that accumulate architecture decisions
Consumer: Personal coaches that track progress · Study assistants with spaced repetition · Health advisors with longitudinal analysis
The pattern: Any scenario where "AI that gets smarter over time" beats "AI that starts from scratch every time."
| Platform | Status | Notes |
|---|---|---|
| Telegram | ✅ Recommended | Full-featured |
| Discord | ✅ | Full-featured |
| 飞书 (Feishu) | ✅ | Native integration |
| 钉钉 (DingTalk) | ✅ | Native integration |
| 企业微信 (WeCom) | ✅ | Native integration |
| REST API | ✅ | For custom frontends |
| 微信 (WeChat) | 🔜 | Coming soon |
| Web UI | 🔜 | Coming soon |
# ~/.leaper/global.yaml (generated by leaper init)
api:
provider: openai # openai / claude / deepseek / ollama / custom
base_url: https://api.openai.com/v1
api_key: sk-xxx
model: gpt-4o
proxy:
http: http://127.0.0.1:10809| Minimum | Recommended | |
|---|---|---|
| Python | 3.10 | 3.12+ |
| RAM | 2 GB | 4 GB+ |
| Network | Required | Stable connection |
| LLM | Any OpenAI-compatible | Claude Opus / GPT-4o |
- Approval system — sensitive operations require human confirmation
- Sandbox isolation — tool execution in sandboxed environment
- Path guard — restrict file system access per agent
- DeepBrain 6-layer memory engine + 4Gate
- Template system (10 CXO + 140 industry)
- Multi-agent single-process architecture
- Telegram / Discord / 飞书 / 钉钉 / 企业微信
- MCP protocol support
- Usage billing + cost optimization
- Web UI management panel
- WeChat integration
- Agent-to-agent collaboration
- Template marketplace (Workshop)
- Benchmark publication
MIT — free for commercial use.
| Project | Description |
|---|---|
| leaper-agent-cn | 🇨🇳 中国版 — 国产 LLM 原生适配，国内渠道深度集成 |
| opc-deepbrain | 🧠 独立的六层记忆引擎 — <2000 行，可嵌入任何框架 |
Deepleaper 跃盟开源
Building AI that learns. 让 AI 越用越聪明。
⭐ Star this repo if you believe AI should remember.