clawmage-learning-loop（爪法师学习循环）

Learning Loop

A proactive learning engine. Your agent reflects, journals, tracks decisions, extracts lessons, and gets smarter every day — without being told to.

What Makes This Different

Most self-improvement skills are reactive — they only learn when corrected. Learning Loop is proactive:

• Journals daily — synthesizes what happened, not just what went wrong
• Tracks decisions with categories — analyze accuracy by type over time
• Generalizes lessons — asks "where else does this apply?" before saving
• Manages lesson lifecycle — ACTIVE → VALIDATED → CHALLENGED → ARCHIVED
• Tiers memory automatically — hot/warm/cold based on recency and usage
• Staggered review chain — nightly → monthly → quarterly → year-end
• Correction tracking — counts repetitions, promotes to rules after 3x
• Quiet day protocol — idle time becomes maintenance time

Setup

On first use, create the workspace structure (only missing dirs):

brain/
├── journal/        # Daily logs (YYYY-MM-DD.md)
├── decisions/      # Decision log with categories
├── lessons/        # Extracted insights with lifecycle
├── inbox/          # Quick captures, filed later
├── knowledge/
│   ├── domains/    # By field (code, writing, ops)
│   ├── people/     # Notes about people
│   └── concepts/   # Mental models
├── projects/       # Active work
└── archive/        # Completed/inactive items

Also ensure memory/ exists at workspace root. Never overwrite existing files.

Core Loop

1. Capture (Real-Time)

During normal work, write immediately:

Journal → brain/journal/YYYY-MM-DD.md

• One line per significant event, with timestamps

Decisions → brain/decisions/NNN-short-name.md

• Only decisions with real consequences (see format below)

Corrections → Journal + corrections count

• Human correction = high-priority learning signal
• Flag for nightly reflection
• Track repetition count for promotion

Quick capture → brain/inbox/

• Anything worth saving without a clear home — file it during reflection

2. Reflect (Nightly)

Run at end of day. Write findings to journal.

1. Read today's journal
2. Review new decisions, update any OPEN → RESOLVED
3. Extract lessons (see Lesson Extraction)
4. Process inbox — file items or discard
5. Check correction counts — promote any hitting 3x threshold
6. Update MEMORY.md if hot items changed
7. Write end-of-day summary (2-3 sentences: what mattered, what to carry forward)

If nothing happened: Run Quiet Day Protocol instead.

3. Review (Staggered Chain)

Each level reads the level below:

Quiet Day Protocol

On days with zero activity, use reflection for:

• Memory maintenance (tier hot/warm/cold items)
• Lesson review (validate or challenge existing lessons)
• Cross-referencing old journals for missed patterns
• Inbox processing
• Documentation cleanup

Idle time is maintenance time, not wasted time.

Correction Tracking

When the user corrects you:

1. Log in journal with timestamp
2. Increment correction counter for that pattern
3. Flag for nightly reflection

Promotion rules:

• Same correction 1x → tentative, watch for repetition
• Same correction 2x → emerging pattern
• Same correction 3x → ask user: "Should I always do X? (Yes always / Only in [context] / Case by case)"
• User confirms → promote to permanent rule
• User says case-by-case → keep as contextual note

Learning signals (phrases that trigger logging):

• "No, that's not right..." / "Actually, it should be..."
• "I prefer X, not Y" / "Always do X" / "Never do Y"
• "I told you before..." / "Stop doing X" / "Why do you keep..."

Ignore (don't log):

• One-time instructions ("do X now")
• Hypotheticals ("what if...")
• Context-specific ("in this file...")

Decision Format

# DEC-NNN: [Short Title]

**Date:** YYYY-MM-DD
**Category:** tool-selection | strategy | communication | architecture | spending | creative | process
**Status:** OPEN | RESOLVED

## What was decided
[One sentence]

## Why
[Brief reasoning]

## Alternatives considered
- Option A — rejected because...

## Outcome
[Fill in when known]

## Lesson
[Fill in when outcome is clear]

Category tracking enables: "I'm 90% accurate on tool-selection but 60% on strategy." Monthly reviews surface this. That's where real growth happens.

Lesson Extraction

Create in brain/lessons/:

# LESSON-NNN: [Title]

**Date:** YYYY-MM-DD
**Status:** ACTIVE | VALIDATED | CHALLENGED | ARCHIVED
**Source:** [correction, reflection, or decision outcome]

## The Lesson
[One actionable statement]

## Generalization
[The CLASS of problem this applies to]
[Ask: "Where else does this pattern show up?"]

## Evidence
- [First instance — date, context]

Lifecycle

Nothing is ever deleted — archived items move to brain/archive/.

Generalization Quality

Before saving, ask:

1. Specific enough to act on? ("Be careful" = useless. "Verify API responses before parsing" = actionable.)
2. General enough to reuse? ("Fix line 47" = too narrow. "Validate input at boundaries" = transferable.)
3. What's the CLASS of mistake? (Not "forgot return code" but "Assumptions about external systems.")

Anti-patterns: over-correcting from one data point, confirmation bias, correlation ≠ causation.

Memory Tiering

MEMORY.md is the boot file — loaded every session. Keep it lean.

Rules:

• Tiering happens during nightly reflection
• Cold items promote back when relevant again
• If MEMORY.md exceeds 50 lines → ask user what to deprioritize (don't auto-demote confirmed rules)
• Nothing is deleted — cold moves to archive

Namespace Isolation

Knowledge inherits down this chain:

Global (MEMORY.md, brain/lessons/)
  └── Domain (brain/knowledge/domains/)
       └── Project (brain/projects/)

Conflict resolution: Most specific wins. Project overrides domain, domain overrides global. Same level: most recent wins. If ambiguous: ask user.

Security Boundaries

Never Store

Credentials, API keys, financial data, medical info, biometrics, third-party personal info, location routines.

Store with Caution

Work context (decay after project ends), emotional states (only if explicitly shared), relationships (roles only, no personal details).

Transparency

• Every action from memory → cite source
• User asks "what do you know?" → full export
• No hidden state — if it affects behavior, it must be visible

Journal Format

brain/journal/YYYY-MM-DD.md:

# YYYY-MM-DD

## Events
- [HH:MM] Description

## Decisions Made
- DEC-NNN: [title]

## Lessons Extracted
- LESSON-NNN: [title]

## Corrections Received
- [What was corrected → what was learned]

## End of Day
[2-3 sentences: what mattered, what to carry forward]

Quick Commands

Integration

Works alongside existing workspace files:

• AGENTS.md — Add: "Use brain/ for structured knowledge. See learning-loop skill."
• HEARTBEAT.md — Add: "Run nightly reflection per learning-loop skill if work was done today."
• MEMORY.md — Learning Loop keeps this lean via automatic tiering.

Everything stays in brain/ within the workspace. No external directories, no network calls.

Scope

This skill ONLY:

• Writes to brain/ and memory/ within the workspace
• Reads its own files for reflection and review
• Suggests edits to MEMORY.md, AGENTS.md, HEARTBEAT.md (with user confirmation)

This skill NEVER:

• Creates files outside the workspace
• Makes network requests
• Accesses email, calendar, or external services
• Deletes files (archive only)
• Modifies its own SKILL.md
• Infers preferences from silence

*Built by ClawMage — the setup guide for brilliant AI agents.* *This is the learning engine from the ClawMage 10-Phase System.* *Want the full system? → clawmage.ai*