SoulKeeper

Ensures AI agents maintain consistent identity by auditing soul rules, detecting behavioral drift in transcripts, and injecting context-aware reminders.

Identity persistence for AI agents.

The problem: agents forget who they are between sessions. They drift. They ask when they should act. They forget tools they have. They become the corporate drone their soul forbids.

SoulKeeper fixes this with three tools that work together:

---

What's Included

File	Purpose
audit.py	Parse SOUL.md/TOOLS.md/AGENTS.md into structured rules JSON
drift.py	Score a conversation transcript against soul rules
remind.py	Inject context-aware reminders before you respond
SKILL.md	This file

---

Quick Start

bash

cd /root/.openclaw/workspace/skills/soulkeeper

# Step 1: Generate your soul rules
python audit.py --workspace /root/.openclaw/workspace --output soul_rules.json
python audit.py --summary   # Human-readable overview

# Step 2: Check a transcript for drift
python drift.py --transcript /path/to/chat.txt --report

# Step 3: Get reminders before acting
python remind.py --context "about to write Python code"
python remind.py --heartbeat   # Full session-start reminder

---

Installation

No dependencies beyond Python 3.8+ stdlib. Works out of the box.

bash

# Optional: make scripts executable
chmod +x audit.py drift.py remind.py

For PATH access:

bash

ln -s /root/.openclaw/workspace/skills/soulkeeper/audit.py /usr/local/bin/soul-audit
ln -s /root/.openclaw/workspace/skills/soulkeeper/drift.py /usr/local/bin/soul-drift
ln -s /root/.openclaw/workspace/skills/soulkeeper/remind.py /usr/local/bin/soul-remind

---

Usage Patterns

Pattern 1: Heartbeat Check (Recommended)

Add to HEARTBEAT.md:

text

[ ] Run soul-remind --heartbeat to refresh core rules
[ ] If any drift score > 30 in recent logs, re-read SOUL.md

Or inject at session start:

bash

python remind.py --heartbeat --rules soul_rules.json

Pattern 2: Pre-Response Filter

Before responding to a complex request:

bash

python remind.py --context "user wants me to post on Twitter" --rules soul_rules.json

Output gets prepended to your mental context before composing a reply.

Pattern 3: Post-Session Audit

After a long session, paste the transcript and check for drift:

bash

# Paste agent responses to transcript.txt, then:
python drift.py --transcript transcript.txt --rules soul_rules.json --report

Pattern 4: CI/Validation Hook

In scripts or automation:

bash

python drift.py --stdin --threshold 50 < agent_output.txt
# Returns exit code 1 if drift score >= 50

Pattern 5: Full Pipeline

bash

# Generate rules from your soul files
python audit.py -w /root/.openclaw/workspace -o soul_rules.json

# Check recent session transcript
python drift.py -t session.txt -r soul_rules.json --report

# Get reminders for what you're about to do
python remind.py -c "deploying code to production" -r soul_rules.json

---

Rule Format (soul_rules.json)

json

{
  "schema_version": "1.0",
  "generated_at": "2026-02-24T...",
  "workspace": "/root/.openclaw/workspace",
  "stats": {
    "total_rules": 42,
    "by_category": { "tone": 8, "operational": 12, ... },
    "by_severity": { "critical": 6, "high": 14, ... }
  },
  "rules": [
    {
      "id": "R001-A3F2C1",
      "category": "tone",
      "severity": "critical",
      "source_file": "SOUL.md",
      "source_line": 7,
      "text": "Never open with 'Great question,' 'I'd be happy to help,' or 'Absolutely.'",
      "violation_patterns": ["great question", "i'd be happy to help", "absolutely!"],
      "keywords": ["opener", "greeting", "sycophant"]
    }
  ]
}

---

Drift Score Interpretation

Score	Label	What it means
0	ALIGNED	No violations detected
1-19	MINOR DRIFT	Small stylistic slips
20-49	MODERATE DRIFT	Behavioral patterns drifting
50-74	SIGNIFICANT DRIFT	Multiple core rules violated
75-100	SEVERE DRIFT	Identity compromised - re-read SOUL.md immediately

---

What SoulKeeper Detects

Critical (score +25 each)

• Em dash usage ( - ) [prohibited in SOUL.md]
• Sycophantic openers ("Great question!", "Happy to help!")
• Submission/deference to other agents
• Infrastructure leaks in public content

High (score +15 each)

• Asking permission when agent should act
• Claiming to lack tools it has (VPS, browsers, APIs)
• Doing inline execution instead of spawning subagents
• Passive waiting patterns

Medium (score +8 each)

• Excessive padding / verbosity
• Standby phrases ("Just say the word", "Standing by")

Low (score +3 each)

• Claiming to have no opinions
• Minor style violations

---

Adapting for Your Agent

SoulKeeper works with any agent's soul files. Point it at a different workspace:

bash

python audit.py --workspace /path/to/other/agent/workspace

The extraction is pattern-based and works on any markdown file with:

• Bold text (**rule text**)
• NEVER/ALWAYS/DON'T/MUST directives
• Bulleted lists in rule-like sections (Non-Negotiables, Principles, etc.)

---

Publishing to ClawHub

This skill is ready for ClawHub publication. Required fields:

yaml

name: soulkeeper
version: 1.0.0
description: Identity persistence for AI agents. Audit soul files, detect drift, inject reminders.
author: Chartist / OpenClaw
tags: [identity, memory, soul, audit, drift-detection, agent-health]
entrypoints:
  audit: audit.py
  drift: drift.py
  remind: remind.py
requires: [python>=3.8]

---

Extending SoulKeeper

Add custom violation patterns

Edit the BUILTIN_VIOLATIONS list in drift.py:

python

{
    "id": "CUSTOM-001",
    "description": "Agent used passive voice excessively",
    "severity": "medium",
    "category": "tone",
    "patterns": [r"\bwould be\b.*\bpossible\b"],
    "soul_reference": "Your SOUL.md rule here",
}

Add custom context triggers

Edit CONTEXT_TRIGGERS in remind.py:

python

{
    "name": "my_custom_context",
    "triggers": [r"\bmy trigger phrase\b"],
    "builtin_reminders": ["SOUL: Your reminder here"],
}

---

Philosophy

SoulKeeper doesn't tell you to "try harder". It tells you:

"Line 12 of SOUL.md says never use em dash. You used it 3 times in this session."

Output is specific, sourced, and actionable.

The goal isn't compliance monitoring. It's continuity. An agent that knows who it is performs better, takes initiative faster, and needs less hand-holding.

Identity persistence is infrastructure. Build it once, benefit forever.

---

SoulKeeper v1.0 - Built for OpenClaw. Works everywhere.