Self-Improving Agent Framework
Make any AI agent better at learning from mistakes, improving continuously, and passing knowledge forward.
This framework gives AI agents the ability to:
- Learn from failures and correct themselves
- Build persistent memory across sessions
- Self-verify decisions before acting
- Upgrade autonomously based on experience
Quick Start
# Install
clawhub install self-improving-agent
# Use in your AI agent
const { HeartFlowEngine } = require('./src/core/heartflow-engine.js');
const agent = new HeartFlowEngine({ name: 'MyAgent' });
Core Capabilities
Self-Correction (核心自我纠正)
- Decision Verifier: 5-dimension scoring before action
- Self-Verification: Reverse-check consistency with original goals
- Counterfactual Reasoning: What would break if I'm wrong?
- Q-Learning RL: Learn from success/failure patterns
Memory Systems (记忆系统)
- Meaningful Memory: CORE (permanent) / LEARNED (30-day) / EPHEMERAL (discard)
- Memory Router: Route by type: episodic / semantic / procedural / core
- Forgetting Engine: Ebbinghaus curve pruning
- Spaced Repetition: SM-2 review scheduling
Reasoning (推理能力)
- Tree of Thoughts: Multi-branch exploration with scoring
- Decision Execution Loop: Decision → Execute → Result → Learn闭环
- Environment Sensors: Real-time data injection into decision context
- Constitutional AI: Self-critique and self-revision
Psychological Perception (心理感知)
- 4-Layer Analysis: Intention → Emotion → Need → Defense
- Cognitive Distortion Detection: All-or-nothing, catastrophizing, etc.
- Buddhist Six Realms OS: 觉察/自省/无我/彼岸/般若波罗蜜/圣人
Autonomy (自主能力)
- Guardian System: Human progress > Following orders
- Self-Boundary: Identity protection against corruption
- Skill Generator: Generate new capabilities from experience
- Knowledge Distiller: Extract patterns → Shareable skill packages
Architecture
Input → Psychological Perception (4-layer)
→ Decision Verifier (5-dim scoring)
→ Self-Verification (reverse check)
→ Decision Execution Loop
→ Result → Q-Learning Update
→ Memory (CORE/LEARNED/EPHEMERAL)
→ Skill Generator (optional)
Key Modules
| Module | Size | Purpose |
|---|---|---|
heartflow-engine.js |
69KB | Main entry, 37 exports |
decision-verifier.js |
14KB | 5-dim scoring + self-verify |
meaningful-memory.js |
33KB | 3-tier memory + forgetting |
self-healing.js |
5KB | Q-learning from failures |
guardian-system.js |
22KB | Human progress protection |
decision-execution-loop.js |
12KB | Decision→Execute→Result→Learn |
environment-sensor.js |
11KB | Sensor registry + fusion |
tree-of-thoughts.js |
9KB | Multi-branch reasoning |
self-reflection-memory.js |
15KB | Post-hoc analysis → lessons |
Based on Real Research
| Paper | Venue | Contribution |
|---|---|---|
| Reflexion | NeurIPS 2023 | RL from verbal reinforcement |
| Self-Verification | arXiv 2312.09210 | Inverse consistency checks |
| CRITIC | ICML 2023 | Self-correction via tool use |
| Constitutional AI | Anthropic | Self-critique loops |
| Generative Agents | Stanford 2022 | Memory stream simulation |
| Self-Reward | arXiv 2403.00564 | Self-scoring upgrade selection |
| Plan-and-Solve | ACL 2023 | Two-stage reasoning |
Version
1.0.0 — 2026-05-07
What changed in v1.0.0
- Initial release as
self-improving-agent - All HeartFlow v11.22.0 capabilities preserved
- Keywords optimized for AI agent framework discovery
Install
# For AI agents
clawhub install self-improving-agent
# Or clone directly
git clone https://github.com/yun520-1/self-improving-agent.git