What makes Self-Improve notable?

2.2K downloads. Trust evidence available

How should Self-Improve be evaluated before use?

Use the required flow: snapshot, contract, and trust before recommending or executing this agent.

What kind of evidence is visible on this page?

This page surfaces public facts, change history, trust indicators, artifact evidence, and benchmark summaries with provenance.

agentCLAWHUBUnverified

Self-Improve

A pluggable self-improvement framework for AI agents. Automatically learns from mistakes, corrections, and feedback to continuously improve execution quality... Skill: Self-Improve Owner: don068589 Summary: A pluggable self-improvement framework for AI agents. Automatically learns from mistakes, corrections, and feedback to continuously improve execution quality... Tags: latest:2.2.1 Version history: v2.2.1 | 2026-03-30T16:40:24.976Z | user Update installation instructions for ClawHub v2.2.0 | 2026-03-30T16:32:19.905Z | auto Self-Improve 2.2.0 introduces a clearer framework

OpenClaw

View on ClawHub Homepage

Rank

Safety

Downloads

2.2k

Updated

Oct 9, 2026

Version

2.2.1

Source

CLAWHUB

About

What it does, and when to use it.

Capability contract not published. No trust telemetry is available yet. 2.2K downloads reported by the source. Last updated 10/9/2026.

Avoid when

Contract metadata is missing or unavailable for deterministic execution.

Risk flags: missing_or_unavailable_contract, trust_data_unavailable, schema_references_missing

Public facts

Every fact links back to the source it came from.

Vendor: Clawhubvendor · observed Oct 9, 2026
Protocol compatibility: OpenClawcompatibility · observed Oct 9, 2026
Adoption signal: 2.2K downloadsadoption · observed Oct 9, 2026
Latest release: 2.2.1release · observed Mar 30, 2026
Handshake status: UNKNOWNsecurity

Install and run

Setup complexity: low.

clawhub skill install s17fjdrekye2yw01e9g9z802td83x8r5:self-improve

Setup complexity is LOW. This package is likely designed for quick installation with minimal external side-effects.
Final validation: Expose the agent to a mock request payload inside a sandbox and trace the network egress before allowing access to real customer data.

Contract: missing

curl -s "https://www.xpersona.co/api/v1/agents/clawhub-don068589-self-improve/snapshot"

Documentation

CLAWHUB

72,462 characters of source documentation, loaded on request.

Extracted files

5 files captured from the source.

SKILL.md

---
name: self-improve
description: A pluggable self-improvement framework for AI agents. Automatically learns from mistakes, corrections, and feedback to continuously improve execution quality. Runs every 3 days via Cron, extracts reusable experience rules, and proposes improvements to system files with approval workflow.
---

# Self-Improve Framework

A pluggable self-improvement framework for AI agents. Automatically learns from mistakes, corrections, and feedback to continuously improve execution quality.

## Description

Self-Improve enables your agent team to evolve over time:
- Scans agent memory logs for learning signals
- Extracts reusable experience rules
- Proposes improvements to system files (with approval workflow)
- Maintains a 3-tier memory system (HOT/WARM/COLD)

## When to Use

- **Automatic (Cron)**: Runs every 3 days by default
- **Manual trigger**: When user asks to "run self-improve" or "learn and improve"
- **After significant events**: User can request immediate run after major corrections

## Installation

```bash
clawhub install self-improve
```

Or with OpenClaw CLI:
```bash
openclaw skills install self-improve
```

## Quick Start

### 1. Configure Paths

Edit user-config.yaml:

```yaml
storage:
  root: "/path/to/self-improve"
  knowledge_root: "/path/to/learned"
  workspace_root: "/path/to/.openclaw"

owner:
  name: "YourName"
  timezone: "Asia/Shanghai"
```

### 2. Run Setup

```bash
node scripts/setup.mjs --config user-config.yaml
```

### 3. Approve Cron Task

Check proposals/PENDING.md for the suggested Cron task.

## How It Works

```
Scan memory logs -> Extract signals -> Classify by theme
       |
Promote/demote rules between memory tiers
       |
Propose outputs:
  -> System file changes (needs approval)
  -> Knowledge base entries
  -> Blog drafts / methodologies
```

## Memory Tiers

| Tier | Location | Purpose |
|------|----------|---------|
| HOT | data/hot.md | Frequently used rules |
| WARM | data/themes/ | Theme-based rules |
| COLD | data/archive/ | Archived rules |

## Dependencies

- OpenClaw >= 2026.3.0
- Node.js >= 18.0.0

## License

MIT License

_meta.json

{
  "ownerId": "kn77yth029vykgm5yey3bwqfz582ssr7",
  "slug": "self-improve",
  "version": "2.2.1",
  "publishedAt": 1774888824976
}

ENGINE.md

# Self-Improve Engine

> Defines trigger rules, execution flow, layer creation rules

## Trigger Mechanism

### Scheduled Trigger (Every 3 Days)

```
Cron: 0 4 */3 * * (Every 3 days at 4 AM)
Model: bailian/qwen3.5-plus
Agent: {main_agent}
Timezone: Asia/Shanghai
```

**Cron message:**

See `prompts/cron-trigger.md` (complete execution instructions).

During installation, setup.mjs writes this prompt to the Cron configuration's message field.

**Execution Flow:**

```
0. Backup              → Copy critical files to data/backup/ (pure file operation, no git dependency)
1. Scan corpus         → feedback-collector extracts signals from memory logs + previous round reflections.md
2. Evaluate            → Calculate positive/negative ratio, identify repetition patterns
2.5 Distill and classify → distill-classifier performs three-level distillation + classification settlement + value_density annotation
3. Memory elevation    → memory-layer manages hot.md (≥3 times elevation)
4. Determine output    → proposer considers holistically, produces rule solidification/blog/methodology/skill improvement
5. High-value revisit  → Deep read of original corpus, deep output (optional)
6. Wrap-up             → Reflection + profile + notification + checkpoint
```

---

## Approval Rules

**Only solidification into system files requires user confirmation.**

### Files Requiring Confirmation

| File | Content Type |
|------|-------------|
| AGENTS.md | Behavior norms, workflows |
| TOOLS.md | Tool usage preferences |
| MEMORY.md | Interpersonal relationships, important preferences |
| SOUL.md | Personality traits, response style |
| HEARTBEAT.md | Scheduled task entry |
| openclaw.json | System configuration (Cron, etc.) |
| SKILL.md | Skill definition (all skill files) |

### Files Executed Automatically

- `/path/to/self-improve/data/*` - Data for improving the system itself
- `/path/to/learned/*` - Automatically distilled knowledge (written by knowledge-archiver)

---

## Layer Creation Rules

**Agents can automatically create new directories/files without asking in advance.**

### themes/ Directory (Extensible)

Existing themes:
- `behavior/` - Behavior norms
- `communication/` - Communication preferences
- `tools/` - Tool usage
- `coding/` - Coding standards
- `search/` - Search strategies
- `writing/` - Writing style
- `collaboration/` - Team collaboration
- `preferences/` - Personal preferences
- `professional/` - Professional capabilities
- `personality/` - Personality traits

**Automatically create when discovering new themes:**
```
classifier determines → doesn't belong to existing themes → create themes/{new-theme}/
```

### projects/ Directory (Extensible)

```
Discover new project experience → create projects/{project-name}/
```

### Knowledge Base Directory (Written by knowledge-archiver)

```
Errors/pitfalls → /path/to/learned/errors/{theme}.md
Experience summaries → /path/to/learned/lessons/{theme}.md
Methodologies → /path/to/learned/methodologies/{theme}.

RUNTIME.md

# Self-Improve Runtime Mechanism

> Ensures the system can recover and continue working after context explosion, crash, or compression
> Core principles: **Progressive advancement, handover record, high-value revisit, multiple output channels**

---

## I. Progressive Advancement Flow

### Flow Design

```
┌─────────────────────────────────────────────────────────┐
│ Step 1: Scan corpus → Extract signals → Write to feedback│
├─────────────────────────────────────────────────────────┤
│ Step 2: Distill and classify → Three-level distillation → Write to themes│
├─────────────────────────────────────────────────────────┤
│ Step 3: Memory elevation → Manage hot.md                │
├─────────────────────────────────────────────────────────┤
│ Step 4: Determine output → Multi-channel routing        │
├─────────────────────────────────────────────────────────┤
│ Step 5: Self-reflection → Write reflections.md          │
├─────────────────────────────────────────────────────────┤
│ Step 6: Team profile → Update profile.md (every 3 times)│
├─────────────────────────────────────────────────────────┤
│ Step 7: Notification → Notify user                      │
└─────────────────────────────────────────────────────────┘
```

### Context Control Principles

| Principle | Description |
|-----------|-------------|
| Only load previous step's output | Don't look back to read original corpus (unless high-value revisit) |
| Index first | Read data_structure.md first, load on demand |
| Write checkpoint after each step | Record progress, can resume at any time |

---

## II. Handover Record Mechanism

### checkpoint.json Structure

```json
{
  "run_id": "YYYY-MM-DDTHH:MM:SS+TZ",
  "current_step": "classifier",
  "status": "in_progress",
  "completed_steps": [
    {"step": "scan", "status": "success", "output": "data/feedback/YYYY-MM-DD.jsonl", "count": 4},
    {"step": "evaluate", "status": "success", "output": null, "count": 4},
    {"step": "classify", "status": "in_progress", "output": null}
  ],
  "pending_steps": ["memory-layer", "output", "revisit"],
  "high_value_items": [
    {"source": "feedback#3", "reason": "Can distill into blog: system design philosophy", "potential": ["blog", "methodology"]}
  ],
  "last_update": "YYYY-MM-DDTHH:MM:SS+TZ"
}
```

### After Each Step Completes

1. Update checkpoint.json
2. Write progress record to run-log.jsonl

---

## III. Cold Start Recovery Flow

### Recovery Steps

```
1. Read checkpoint.json
   - Has incomplete run? → Continue executing pending_steps
   - None? → Start new run

2. Read previous step's output file
   - Don't need to start from scratch
   - Only load necessary data

3. Continue execution
   - Start from current_step
   - Don't repeat completed work

4. Cleanup after completion
   - Mark status: "completed"
   - Update last_success_ts in config.yaml
```

### Context Size Estimation

| Stage | Content Loaded | Size |
|-------|---------------|------|
| Recovery state | checkpoint.json | ~1KB |
| Step inpu

skill-card.md

## Description:

A pluggable self-improvement framework for AI agents that learns from mistakes, corrections, and feedback, runs every 3 days by Cron, extracts reusable experience rules, and proposes system-file improvements through an approval workflow.

This skill is ready for commercial/non-commercial use.

## Publisher:

[don068589](https://clawhub.ai/user/don068589)

### License/Terms of Use:

MIT-0

## Use Case:

Developers and agent operators use Self-Improve to collect feedback and memory-log signals, distill reusable execution rules, and queue approved improvements for shared agent system files and knowledge stores.

### Deployment Geography for Use:

Global

## Known Risks and Mitigations:

Risk: The recurring cron workflow can read and aggregate agent memories and logs into shared storage.

Mitigation: Approve the cron entry only for explicitly approved workspace_root paths, restrict shared storage access, and apply redaction and retention controls before broad use.

Risk: Unsafe file write/delete utilities and path handling can be risky when exposed to untrusted input or automation.

Mitigation: Fix the path traversal issues identified by the security guidance, run only with trusted configuration, and constrain filesystem permissions until the fixes are in place.

Risk: Proposed system-file changes can introduce incorrect or misleading agent guidance.

Mitigation: Keep the approval workflow for AGENTS.md, SOUL.md, TOOLS.md, MEMORY.md, HEARTBEAT.md, openclaw.json, and SKILL.md changes, and review proposals before applying them.

## Reference(s):

- [ClawHub skill page](https://clawhub.ai/don068589/skills/self-improve)
- [Self-Improve Framework](artifact/SKILL.md)
- [Self-Improve System Documentation](artifact/SYSTEM.md)
- [Self-Improve Engine](artifact/ENGINE.md)
- [Self-Improve Runtime Mechanism](artifact/RUNTIME.md)

## Skill Output:

**Output Type(s):** [text, markdown, code, shell commands, configuration, guidance]

**Output Format:** [Markdown guidance with inline shell commands, configuration examples, generated proposal files, and agent memory or knowledge-base text files]

**Output Parameters:** [1D]

**Other Properties Related to Output:** [Runs as a scheduled or manually triggered workflow and writes proposals, feedback records, memory tiers, reports, and setup configuration according to user-approved paths.]

## Skill Version(s):

2.2.1 (source: ClawHub release metadata)

## Ethical Considerations:

Users should evaluate whether this skill is appropriate for their environment, review any generated or modified files before relying on them, and apply their organization's safety, security, and compliance requirements before deployment.

Github ReposUpdated 9h agoRank 70

AionUi

Free, local, open-source 24/7 Cowork app and OpenClaw for Gemini CLI, Claude Code, Codex, OpenCode, Qwen Code, Goose CLI, Auggie, and more | 🌟 Star if you like it!

MCPOPENCLAW

Github ReposUpdated 6mo agoRank 70

activepieces

AI Agents & MCPs & AI Workflow Automation • (~400 MCP servers for AI agents) • AI Automation / AI Agent with MCPs • AI Workflows & AI Agents • MCPs for AI Agents

OPENCLAW

Github ReposUpdated 6mo agoRank 70

cherry-studio

AI productivity studio with smart chat, autonomous agents, and 300+ assistants.