Description

Experience distillation engine that turns raw daily memory logs into compounding intelligence. Extracts patterns, generates briefings, tracks growth metrics,...

README (SKILL.md)

CompoundMind v0.1

Name: CompoundMind
Author: cassh100k

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

Distills memory logs into structured lessons, decisions, skill updates, relationships, and facts
Indexes everything into a searchable SQLite database weighted by recency and importance
Briefs you before any task with targeted lessons from past experience
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

Command	What it does
`sync`	Distill all new memory files + rebuild index
`distill`	Extract structured knowledge from memory files
`rebuild`	Rebuild the SQLite wisdom index
`search \x3Cquery>`	Search accumulated wisdom
`brief \x3Ctask>`	Pre-session briefing for a specific task
`report`	Generate growth report with LLM narrative
`mistakes`	Show repeated mistake patterns
`stats`	Index statistics

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:

After each session - run sync (or schedule via cron)
Before each task - run brief "task description"
Weekly - run report to see growth trajectory
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

Usage Guidance

This skill largely does what it says, but there are important privacy/incoherence issues to consider before installing or running it: - The package includes many pre-populated experience files with wallet addresses, env mentions (e.g., POLYMARKET_KEY) and other operational details. Treat those files as sensitive data and inspect/delete them before running. - Although SKILL.md claims 'runs entirely local', the distiller and optional LLM briefing paths call Anthropic (Claude Haiku). If you provide ANTHROPIC_API_KEY (or COMPOUND_MIND_LLM_KEY) the skill will send memory contents to a remote LLM. If you do not want that, do not set an LLM key and avoid the --llm flags (or audit distill.py/build_report_llm/build_briefing_llm to ensure no network calls occur). - Review distill.py to confirm exactly what text is sent to the LLM (full files vs redacted summaries). If the code sends raw memory files, remove or redact secrets from your memory store first. - If you want to use the skill: run it in an isolated sandbox, back up and/or remove sensitive memory files from /root/.openclaw/workspace/memory, and consider deleting the included data/experiences if they contain real operational secrets. - Prefer the rule-based (local) synthesis paths if you need local-only operation; verify those code paths never import anthropic or perform network I/O. If you want, I can (1) point out the exact lines in distill.py/brief.py that call the Anthropic API, (2) show where the code reads memory files so you can confirm what's transmitted, or (3) suggest a minimal patch to disable remote LLM calls and redact secrets before distillation.

Capability Analysis

Type: OpenClaw Skill Name: compound-mind Version: 0.1.0 The skill is classified as suspicious due to its handling and storage of sensitive information, and the inherent prompt injection risk in its LLM interactions. The `distill.py` module explicitly extracts and stores 'Facts' such as API keys, wallet addresses, and configuration values in plain text JSON files within the skill's local `data/experiences/` directory (e.g., `data/experiences/0bfdeb973bbe.json` and `data/experiences/2e40b3400fb0.json`). While this is part of its stated purpose to build an 'experience index', storing such sensitive data unencrypted locally presents a significant data security vulnerability. Additionally, the `brief.py` and `growth.py` modules construct LLM prompts using content derived from the agent's memory logs, which could be manipulated by an attacker to achieve prompt injection against the LLM, even with defensive instructions in the prompt. There is no evidence of intentional exfiltration or other malicious behavior, but these vulnerabilities elevate the classification to suspicious.

Capability Assessment

ℹ Purpose & Capability

The code (compound_mind.py, distill.py, brief.py, index.py, growth.py) and SKILL.md align with the described purpose (distill memory logs, build a wisdom index, generate briefs, track growth). However the registry metadata you provided says 'Required env vars: none / Primary credential: none' while clawpkg.yaml and the code clearly expect an Anthropic API key (ANTHROPIC_API_KEY or COMPOUND_MIND_LLM_KEY) for LLM-based distillation/briefing. That metadata mismatch is an incoherence to surface.

⚠ Instruction Scope

SKILL.md and the code instruct the skill to read memory files from /root/.openclaw/workspace/memory/ and distill them using Claude Haiku (Anthropic). That means potentially sensitive local memory contents (wallet addresses, API keys, configs) will be sent to an external LLM provider when the LLM path is used. SKILL.md's claim 'Runs entirely local with minimal API calls' is misleading because distillation/briefing uses remote LLMs unless you explicitly avoid the LLM code paths. The packaged data/experiences already contain wallet addresses and env mentions, demonstrating the type of sensitive data that could be processed/transmitted.

ℹ Install Mechanism

There is no install spec (instruction-only style), so nothing is downloaded at install time — lower install risk. clawpkg.yaml lists 'anthropic' as a required Python package but no automated installer is provided. Note: the package contains many pre-populated data/experiences files (with potentially sensitive, personal operational data) which increases privacy exposure even before any runtime actions.

⚠ Credentials

The code expects an LLM API key (checks COMPOUND_MIND_LLM_KEY or ANTHROPIC_API_KEY) even though registry metadata declared no required env vars. Requesting an Anthropic key is proportionate to optional LLM features, but it becomes a privacy/credential risk because the skill will read user memory files and could send secrets to the external service. There are also embedded sensitive items in packaged experience files (wallet addresses, an inline mention of POLYMARKET_KEY stored in a script) — the skill does not request unrelated third-party credentials, but it will process and surface secrets found in the memory data.

✓ Persistence & Privilege

always:false and no special OS restrictions; the skill does not declare 'always: true' and does not appear to modify other skills or system-wide agent settings. It persists data only in its own data/ subdirectory (briefs, experiences, wisdom.db). It suggests a cron example (user-installable) and a heartbeat hook in clawpkg.yaml, but these are optional and not grant-of-privilege features on their own.

Version History

v0.1.0

v0.1.0: Experience distillation - turns daily logs into compounding intelligence. Briefings, growth, patterns.

Metadata

Slug compound-mind

Version 0.1.0

License —

All-time Installs 1

Active Installs 1

Total Versions 1

Frequently Asked Questions

What is CompoundMind?

Experience distillation engine that turns raw daily memory logs into compounding intelligence. Extracts patterns, generates briefings, tracks growth metrics,... It is an AI Agent Skill for Claude Code / OpenClaw, with 441 downloads so far.

How do I install CompoundMind?

Run "/install compound-mind" in the OpenClaw or Claude Code chat to install it in one step — no extra setup required.

Is CompoundMind free?

Yes, CompoundMind is completely free (open-source). You can download, install and use it at no cost.

Which platforms does CompoundMind support?

CompoundMind is cross-platform and runs anywhere OpenClaw / Claude Code is available (cross-platform).

Who created CompoundMind?

It is built and maintained by Cassh (@cassh100k); the current version is v0.1.0.

More Skills

CompoundMind