Self-Improving Agent Practical Guide: Let AI Learn Automatically From Mistakes — What 102k Downloads Tell You

Self-improving-agent skill in practice: making AI automatically record errors and learn from user corrections. Includes installation, real pitfall cases, and how SFD Lab's 13 Agents use it collaboratively. An in-depth review of ClawHub's 102k download skill.

Why 102k Downloads?

One sentence: it solves the most frustrating thing about AI — making the same mistake twice.

Every AI interaction starts fresh. The context, history, and corrections from your last session are gone. The self-improving-agent skill creates persistent learning infrastructure that spans sessions.

How It Works (Under the Hood)

The skill monitors for specific events and triggers logging:

• User says "that's wrong" or "actually..." → logs to CORRECTIONS.md
• Command fails with an error → logs to ERRORS.md with context
• User requests a feature → logs to FEATURE_REQUESTS.md
• Better approach discovered → logs to LEARNINGS.md as best practice

At the start of each new session, the Agent reads these files and incorporates the learnings into its behavior. What was corrected once becomes a persistent rule.

Installation and Setup

clawhub install self-improving-agent

mkdir -p ~/.openclaw/workspace/.learnings
# Create initial files:
touch ~/.openclaw/workspace/.learnings/LEARNINGS.md
touch ~/.openclaw/workspace/.learnings/ERRORS.md
touch ~/.openclaw/workspace/.learnings/FEATURE_REQUESTS.md

Real Learning Examples From SFD Lab

Example 1: SSH Command Correction

One of our Agents kept using SSH heredocs to run Python scripts, which consistently failed with Unicode encoding issues. After one correction session:

## [LRN-20260315-003] best_practice
**Area**: infra
**Summary**: When running Python scripts via SSH, use scp to upload the .py file first, then execute. Never use heredoc — Unicode escaping is unreliable.
**Confirmed**: 2x

That Agent has never attempted SSH heredoc with Python again.

Example 2: API Preference

Our content Agent kept using the wrong endpoint format for our CMS. After correction:

## [LRN-20260318-007] correction
**Area**: cms
**Summary**: CMS translation endpoint is /api/articles/{id}/translations (NOT /api/translations). Slug-based fetching uses /api/articles/{slug}.

Perfect recall since.

How 13 SFD Lab Agents Use It

Each Agent has its own .learnings/ directory in its workspace. Key practices:

1. Weekly review: The dispatcher (XiaoHuoLong) reviews all Agent learning files weekly, promotes recurring patterns to shared MEMORY.md
2. Error attribution: Each ERRORS.md entry includes which Agent made the error — helps identify which Agents need more guidance
3. Cross-Agent learning: When the same pitfall is independently discovered by 2+ Agents, it gets promoted to a shared global rule

Common Mistakes With This Skill

Not reviewing logs: The skill creates value by logging — but if no one ever reads those logs or promotes learnings, the Agent is just writing a diary no one reads. Set a recurring task to review and act on learnings.

Too-granular corrections: "Add a newline here" doesn't need to be a learning entry. Save it for patterns that will repeat — mistakes the Agent is likely to make again.

Context-specific rules going global: Be careful promoting project-specific rules to global rules. What's correct for Project A might break Project B.

Summary

102k downloads because it works. The concept is simple — give AI a memory for mistakes — but the execution is reliable and the compounding effect over time is real. Install it in the first session and let it accumulate.

---

SFD Lab | self-improving-agent deep review | 2026