cognitive-distillation

name	cognitive-distillation
description	Four-layer cognitive distillation system. Use when performing daily reviews, recording decisions, extracting patterns, or updating principles. Transforms raw data into durable wisdom through systematic compression: Raw → Episodic → Semantic → Principles.
platform	both
scope	agent:wendy

Cognitive Distillation

Overview

A four-layer system that transforms raw information into durable principles through progressive compression. Each layer distills the one below it.

L4: PRINCIPLES.md        ← 原子原则 (每周 review)
L3: memory/patterns.md   ← 认知模式 (每日蒸馏)
L2: memory/decisions/    ← 决策日志 (实时记录)
L1: memory/*.md + Obsidian ← 原始数据 (自动积累)

The Four Layers

Layer 1: Raw Data (原始数据)

Files: memory/YYYY-MM-DD.md, Obsidian briefings, conversation logs Update: Continuous / automatic Purpose: Capture everything. Don't filter at this stage.

Already handled by OpenClaw's memory flush + cron briefings.

Layer 2: Episodic Memory (事件 + 决策日志)

Files: memory/decisions/YYYY-MM-DD-<topic>.md Update: Whenever a significant decision is made Purpose: Record WHY, not just WHAT.

Decision log template:

# Decision: <title>
Date: YYYY-MM-DD
Context: <what situation triggered this decision>
Options:
  1. <option A> — pros / cons
  2. <option B> — pros / cons
  3. <option C> — pros / cons
Selected: <which option>
Reason: <why this one>
Expected outcome: <what we expect to happen>
Actual outcome: <filled in later>
Lesson: <filled in later>

When to create a decision log:

• Architecture or design choices
• Trading decisions (entry/exit thesis)
• Tool/framework selections
• Strategy pivots
• Any choice Park explicitly deliberates on

Layer 3: Semantic Memory (认知模式)

File: memory/patterns.md Update: Daily distillation cron (22:00) Purpose: Extract recurring patterns from L1 + L2.

Format:

# Patterns

## Park's Working Style
- <observation> (first noted: YYYY-MM-DD, confirmed: N times)

## Market Patterns
- <pattern> (evidence: ...)

## System/Engineering Patterns
- <pattern> (evidence: ...)

## Anti-Patterns (things that failed)
- <what> → <why it failed> (date)

Rules:

• Only add patterns observed 2+ times
• Include counter-evidence when it exists
• Mark confidence: 🟢 strong / 🟡 emerging / 🔴 challenged
• Remove or downgrade patterns that get contradicted

Layer 4: Core Principles (原子原则)

File: PRINCIPLES.md (workspace root) Update: Weekly review (Sunday 20:00) Purpose: The most compressed, highest-value knowledge.

Format:

# Principles

## Investing
- **<principle>** — <one-line explanation> [weight: 1-10] [last validated: date]

## Building
- **<principle>** — <one-line explanation> [weight: 1-10] [last validated: date]

## Meta / Thinking
- **<principle>** — <one-line explanation> [weight: 1-10] [last validated: date]

Rules:

• Max 30 principles total (forces prioritization)
• Each must be actionable (not platitude)
• Weight reflects confidence (1=hypothesis, 10=proven)
• Weekly review: validate, challenge, merge, or remove
• New principle needs 3+ supporting observations from L3

Distillation Processes

Daily Distillation (每日蒸馏, 22:00)

1. Read today's L1 data (memory/YYYY-MM-DD.md, conversation history)
2. Read any new L2 decision logs
3. Ask three questions:
   • What did we learn today that we didn't know yesterday?
   • Was any existing pattern (L3) confirmed or challenged?
   • Did any decision (L2) produce an unexpected outcome?
4. Update memory/patterns.md if warranted
5. Post summary to 🧠 Meta topic for Park to review

Weekly Principle Review (每周原则审查, Sunday 20:00)

1. Read all L3 pattern updates from the past week
2. Read all L2 decision outcomes
3. For each existing principle in L4:
   • Was it validated this week? → update date
   • Was it challenged? → note evidence, consider downgrade
   • Is it redundant with another? → merge
4. Any new pattern strong enough (3+ observations) to become a principle?
5. Post review to 🧠 Meta topic

Heartbeat Check (每次 heartbeat)

Quick scan: is there an unlogged significant decision from recent conversation? If yes: create L2 decision log. If no: pass.

Integration with OpenClaw

Cron jobs needed:

• Daily distillation: 0 22 * * * → isolated agentTurn
• Weekly review: 0 20 * * 0 → isolated agentTurn

Files to create on first run:

mkdir -p memory/decisions
touch memory/patterns.md
touch PRINCIPLES.md

Key Principles of This Skill

• Compression over accumulation — The goal is less data, more wisdom
• Evidence-based — No principle without supporting observations
• Mutable — Principles can be wrong. Downgrade or remove when contradicted
• Human-in-the-loop — Park reviews weekly; the system proposes, Park decides
• Model-upgrade amplified — More data + better model = deeper distillation