Compound Mind

CompoundMind v0.1

Makes agents permanently smarter. Each interaction compounds.

The problem: agents start from zero every session. Reading files helps, but raw logs are bulk. Real intelligence requires distillation - extracting what matters and making it instantly searchable.

What It Does

1. Distills memory logs into structured lessons, decisions, skill updates, relationships, and facts
2. Indexes everything into a searchable SQLite database weighted by recency and importance
3. Briefs you before any task with targeted lessons from past experience
4. Tracks growth over time - are you getting smarter or repeating mistakes?

Quick Start

cd /root/.openclaw/workspace/compound-mind

# Full pipeline (distill all memory files + build index)
python compound_mind.py sync

# Search your accumulated wisdom
python compound_mind.py search "Polymarket order types"
python compound_mind.py search "git mistakes" --category lesson
python compound_mind.py search "Chartist" --category relationship

# Pre-session briefing before a task
python compound_mind.py brief "trade on Polymarket BTC markets"
python compound_mind.py brief "post content on X"
python compound_mind.py brief "debug a Python cron job"

# Growth report
python compound_mind.py report

# Find repeated mistakes
python compound_mind.py mistakes

# Stats
python compound_mind.py stats

Commands

File Structure

compound-mind/
  compound_mind.py    - Main CLI
  distill.py          - Experience distiller (uses Claude Haiku)
  index.py            - SQLite FTS wisdom index
  brief.py            - Pre-session briefing generator
  growth.py           - Growth tracker and report generator
  data/
    experiences/      - Per-source distilled experience JSON files
    wisdom.db         - SQLite FTS database
    growth.json       - Growth tracking state
    briefs/           - Saved pre-session briefs
    distill_state.json - Tracks which files have been processed

How It Works

Distiller

Reads each memory file through Claude Haiku. Extracts:

• Lessons - what worked, what failed, tagged by domain and outcome
• Decisions - context + action + outcome triplets
• Skill updates - evidence of capability improvement over time
• Relationships - how each person communicates, what they prefer
• Facts - specific knowledge worth keeping (wallet addresses, API patterns, config values)

Files are hash-tracked - re-runs only process changed files.

Wisdom Index

SQLite with FTS5 full-text search. Each entry scored by:

• FTS relevance (BM25)
• Recency (exponential decay, 30-day half-life)
• Importance (1-5 score assigned by distiller)

Pre-Session Briefing

Given a task description, detects relevant domains, pulls top wisdom, synthesizes a sharp briefing via Claude Haiku. Covers:

• Critical lessons to remember
• Past failures to avoid
• Key facts and configs needed

Growth Tracker

Analyzes all experience files to compute:

• Lesson positive/negative ratios by domain
• Decision quality rates
• Repeated mistake patterns (same negative lesson appearing across multiple dates)
• Composite growth score (0-100)

Integration with Agent Workflow

Ideal usage pattern:

1. After each session - run sync (or schedule via cron)
2. Before each task - run brief "task description"
3. Weekly - run report to see growth trajectory
4. When stuck - run search "relevant topic" to surface past experience

Cron Example (daily distillation)

0 3 * * * cd /root/.openclaw/workspace/compound-mind && python compound_mind.py sync --since $(date -d "2 days ago" +%Y-%m-%d) >> /tmp/compound-mind.log 2>&1

Dependencies

• Python 3.10+
• anthropic Python SDK (for distillation and briefing)
• SQLite3 (stdlib)
• Memory files at /root/.openclaw/workspace/memory/

No external databases. No vector embeddings. Runs entirely local with minimal API calls.

Design Principles

• Incremental - only re-processes changed files
• Cheap - uses Claude Haiku for extraction (low cost per memory file)
• Fast - SQLite FTS5 for sub-second search
• Honest - growth tracking measures actual quality, not just quantity
• Composable - each module works standalone or as part of the pipeline