Procedural Memory

Inspired by Hermes Agent's procedural memory system. When a workflow works, save it as a reusable skill. Triggered when: (1) user says 'remember this workflo...

Security Scan

The skill's behavior largely matches a 'save successful workflows to disk' feature, but there are several surprising or under-specified behaviors (automatic creation without asking in some cases, autonomous SOP that mentions external backups) that don't line up with what's declared and that pose privacy/consent risks.

Purpose & Capability

The skill's declared purpose (turning repeatable successful workflows into reusable SKILL.md files) matches the instructions to scan conversations, extract patterns, and write SKILL.md files under ~/.openclaw/workspace/skills and logs under ~/self-improving. Writing files and creating directories is coherent for this purpose. Minor mismatch: the autonomous SOP mentions backing up 'GitHub and E盘' (E drive), which implies access to external services/storage not otherwise requested or explained.

Instruction Scope

SKILL.md explicitly instructs the agent to scan conversation history to identify workflows and persist derived summaries to disk. That is within the stated goal but is a clear privacy surface (sensitive conversation content could be summarized and stored). It also contains an explicit rule: auto-create a skill without asking if the same workflow succeeds 3+ times — this creates persistent artifacts on disk without user confirmation in some cases, which may be surprising to users. The instructions reference local shell commands (mkdir, ls, head) and specific paths, so they assume filesystem write/read privileges.

Install Mechanism

Instruction-only skill with no install spec and no code files. No binaries, downloads, or installs are requested, so there is minimal install-time risk.

Credentials

The skill declares no required env vars or credentials, which is consistent with purely local file operations. However, the autonomous SOP's allowed tasks include backing up 'GitHub and E盘' and 'collecting skill information' which could require network access and credentials; those credentials are not declared nor are instructions provided for obtaining them. This is a proportionality mismatch and an under-specification of what external access would need.

Persistence & Privilege

The skill persists summaries and SKILL.md files to the user's home directory. Persisting user-derived data is expected for a memory system, but the combination of: (a) automatic creation of skill files when a pattern repeats 3+ times (without asking), and (b) an autonomous-operation SOP that authorizes periodic autonomous work (after 30+ minutes of inactivity) raises consent and scope concerns. The skill is not 'always: true', but the SOP could enable background autonomous actions if the agent adopts it.

What to consider before installing

This skill is plausible for saving reusable workflows, but review and consider the following before installing: - Consent & auto-creation: SKILL.md says it will auto-create a skill (write files) if the same workflow succeeds 3+ times without asking. If you do not want the agent to create persistent artifacts without explicit approval, edit/remove that auto-create rule. - Privacy: the skill scans conversation history and writes derived summaries to disk (~/.openclaw/workspace/skills and ~/self-improving). Avoid sharing secrets/sensitive data in conversations or change the skill to redact or exclude private channels before saving. - Autonomous SOP mismatch: autonomous-operation-sop.md allows tasks like 'backup GitHub' and scanning for skill info. Backing up external services typically requires credentials and network access — the skill declares no creds. If you permit autonomous background tasks, explicitly constrain them and require approval for any network/external backups. - Verify filesystem impact: ensure your environment allows the agent to write only to intended paths and that those paths don't contain other sensitive data. Consider sandboxing the skill's workspace or adjusting paths to a controlled directory. - If you need higher assurance, ask the skill author to clarify: (1) whether auto-create truly runs without explicit user opt-in, (2) how autonomous SOP is invoked and whether it will perform network operations by default, and (3) how saved SKILL.md/log content is sanitized to avoid accidental leakage of secrets. Given the above gaps and consent/privacy implications, treat this skill as suspicious until you confirm or constrain its automatic behaviors.

Procedural Memory — Automatic Workflow Memory

"When a workflow works, save it as a reusable skill."

Inspired by Hermes Agent's core feature: an agent that learns from experience by saving successful workflows as reusable skills.

---

Directory Prerequisites

mkdir -p ~/self-improving
mkdir -p ~/.openclaw/workspace/skills

The log file ~/self-improving/procedural-memory-log.md will be created automatically if missing.

---

Trigger Conditions (Priority Order)

🔴 High Priority — Explicit User Request

• "记住这个workflow" / "remember this workflow"
• "以后遇到类似就这样做" / "do this next time"
• "这个方法很好，以后也这样用" / "this works, use it again"
• "这个技巧值得记住" / "this is worth remembering"

🟡 Medium Priority — Proactive Review

• End of day (or user says "review")
• After a complex task completes successfully

🟢 Low Priority — Pattern Recognition

• Same workflow applied successfully 2+ times
• Recurring solutions discovered

---

Workflow

Step 1 — Evaluate Worthiness

Must meet ALL:

• ✅ Takes 3+ steps to complete
• ✅ Likely applicable to future conversations
• ✅ Not one-time / context-specific
• ✅ User would benefit

Auto-trigger: If the same workflow succeeds 3+ times, auto-create skill WITHOUT asking (记录到日志即可)。

Otherwise: ask user "我可以把这个 workflow 存为 skill 吗？"

If user confirms → continue. Otherwise → silent skip.

Step 2 — Check for Conflicts

ls ~/.openclaw/workspace/skills/ | grep <keyword>

• Same name exists → log "already exists, skip"
• Similar function exists → consider merge or append

Step 3 — Extract Pattern

Record:

**Skill Name**: [verb]-[problem] (kebab-case)
**Trigger**: [when this would be used]
**Problem Solved**: [what issue this solves]
**Steps**: [numbered list]
**Key Insight**: [why this works]
**Limitations**: [when NOT to use]

Step 4 — Create Skill File

Create ~/.openclaw/workspace/skills/[skill-name]/SKILL.md:

---
name: [skill-name]
description: "[one-line description of when to use]
---
# [Skill Name]

## What This Solves
[trigger conditions and problem]

## When to Use
- [specific scenario 1]
- [specific scenario 2]

## How To Execute
[Step 1]
[Step 2]
[Step 3...]

## Key Insight
[why this is effective]

## Limitations
[when this won't work]

Step 5 — Validate

ls ~/.openclaw/workspace/skills/[skill-name]/SKILL.md
head -20 ~/.openclaw/workspace/skills/[skill-name]/SKILL.md

Recreate if corrupted.

Step 6 — Log

Append to ~/self-improving/procedural-memory-log.md:

## YYYY-MM-DD

### [skill-name]
- **Trigger**: [what triggered this]
- **Problem solved**: [what was solved]
- **Key insight**: [the key learning]
- **Skill file**: `~/.openclaw/workspace/skills/[skill-name]/SKILL.md`

Step 7 — Notify User

Tell user the skill was created and ask if they want to enable auto-execution:

"已记住这个 workflow，存为 skill/[skill-name]。是否需要我下次遇到类似情况时自动执行？"

Default behavior: Do NOT auto-execute without user confirmation.

---

Skill Loading

When new conversation context matches a saved skill's trigger:

1. Load the skill's SKILL.md
2. Ask user: "我之前学过这个，要我用这个方法吗？"
3. Only execute AFTER user confirms
4. Never auto-execute without explicit user approval

Privacy Consideration

• This skill scans conversation history to identify worth-saving workflows
• Derived content is persisted to disk (~/.openclaw/workspace/skills/ and ~/self-improving/)
• Sensitive conversation data may be captured — user is responsible for what they share in conversations
• Log files contain pattern summaries, not raw conversation, but could still contain sensitive references
• Users should be aware of what they discuss with the agent

---

End-of-Day Review Prompt

When user says "复盘" or at end of day:

"今天有什么 workflow 值得我记住吗？"

Scan today's conversation for:

• Complex but successful tasks
• Repeatedly used techniques
• Reusable solutions

---

References

• Skill Creator Guide: ../skill-creator/SKILL.md
• Self-Improving Memory: ~/self-improving/memory.md
• Procedural Memory Log: ~/self-improving/procedural-memory-log.md