DEV Community

Tom Lee
Tom Lee

Posted on • Originally published at blog.clawsouls.ai

Building Safe Agents with Long-Term Memory: SoulScan, Persona Engine & Swarm Memory

#soulclaw #soulscan #persona #swarm

Claude Dispatch Validated the Market. Now Let's Talk About Safety.

Anthropic recently launched Claude Dispatch — a phone-to-desktop agent workflow. This validates what the OpenClaw community has been building for months: AI agents that work autonomously on your behalf.

But there's a gap nobody talks about: How do you keep an autonomous agent safe?

When your agent runs 24/7, handles sensitive data, and has tool access, three problems emerge:

  1. Soul file tampering — Someone (or something) modifies your agent's personality definition
  2. Persona drift — The agent gradually deviates from its defined character over long conversations
  3. Memory fragmentation — Multiple agents can't share what they've learned

SoulClaw v2026.3.21 addresses all three.

1. SoulScan: Inline Security Scanning

SoulScan is a 4-stage security pipeline that scans soul files (SOUL.md, soul.json) for:

  • Prompt injection — Hidden instructions in personality definitions
  • Data exfiltration — Patterns that leak sensitive information
  • Harmful content — 58+ security rules
  • Schema violations — Structural issues

What's New: Inline Scanning

Previously, SoulScan only ran manually or during soul package publishing. Now it runs automatically after every agent turn (rate-limited to once per 5 minutes):

Agent Turn → Response → [fire-and-forget] SoulScan checks workspace
                                          ↓
                                    Score < threshold?
                                          ↓
                                    ⚠️ Warning logged
Enter fullscreen mode Exit fullscreen mode

No configuration needed. If your workspace has a SOUL.md, it's protected.

CLI Usage 

# Scan your workspace
soulclaw soulscan

# CI/CD pipeline
soulclaw soulscan --json --min-score 70
Enter fullscreen mode Exit fullscreen mode

2. Persona Engine: Drift Detection

Here's a subtle problem: even with a perfect SOUL.md, your agent's personality drifts over long conversations. The LLM gradually shifts away from the defined persona — especially after many tool calls, context compactions, or topic changes.

The Persona Engine monitors this drift and catches it before it becomes a problem.

How It Works

  1. Parse — SOUL.md is parsed into structured rules (tone, style, principles, boundaries)
  2. Detect — Every N responses, the last assistant message is scored against these rules
  3. Enforce — If drift exceeds a threshold, a correction is injected + notification sent

Detection methods:

  • Ollama LLM (default) — Sends the response + persona rules to a local model for scoring
  • Keyword matching (fallback) — Checks for tone/style keyword alignment when Ollama is unavailable

Opt-In Configuration

Drift detection is off by default — enable it when you're ready:

# Enable
soulclaw persona config --enable

# Customize
soulclaw persona config --interval 3 --threshold 0.4

# Check manually
soulclaw persona check --text "Your agent's response here"

# View drift history
soulclaw persona metrics
Enter fullscreen mode Exit fullscreen mode

Real-Time Notifications

When drift is detected, you get notified via your configured messaging channel:

⚠️ Persona Drift WARNING
Score: 0.450 (method: keyword)
Session: agent:main:telegram:12345
Action: reminder
Enter fullscreen mode Exit fullscreen mode

No more wondering if your agent is "still in character."

3. Swarm Memory: Multi-Agent Sync

If you run agents across multiple machines — or want multiple agents to share knowledge — Swarm Memory provides Git-based memory synchronization.

The Problem

Agent A learns something important on your Mac. Agent B, running on your server, has no idea. They're working in isolation, duplicating effort, making contradictory decisions.

The Solution 

Agent A (Mac)                    Agent B (Server)
    │                                │
    ├── MEMORY.md                    ├── MEMORY.md
    ├── memory/*.md                  ├── memory/*.md
    │                                │
    └─── swarm sync ──→ Git Repo ←── swarm sync ───┘
Enter fullscreen mode Exit fullscreen mode

Setup 

# Initialize
soulclaw swarm init --remote git@github.com:user/swarm-memory.git

# Check status
soulclaw swarm status

# Force sync
soulclaw swarm sync
Enter fullscreen mode Exit fullscreen mode

LLM-Powered Conflict Resolution

When both agents modify the same memory file, traditional merge fails. Swarm Memory offers LLM semantic merge:

# Sync with intelligent merge
soulclaw swarm sync --llm-merge

# Resolve specific conflicts
soulclaw swarm resolve MEMORY.md --llm
Enter fullscreen mode Exit fullscreen mode

The LLM reads both versions, preserves unique information from each, removes duplicates, and produces a clean merged result. Falls back to "ours" strategy if Ollama is unavailable.

Other resolution options:

soulclaw swarm resolve --ours     # Keep our version
soulclaw swarm resolve --theirs   # Keep their version
soulclaw swarm resolve --manual   # Edit manually
Enter fullscreen mode Exit fullscreen mode

Why This Matters

These aren't academic features. They solve real problems:

Problem Before After
Soul file tampered Agent silently compromised SoulScan catches it within 5 minutes
Agent drifts from character Nobody notices until it's bad Persona Engine alerts you at threshold
Multi-agent knowledge Each agent starts from zero Swarm Memory shares across all agents

Free, Open Source, Built on OpenClaw

All of this is:

  • Free — No subscription required
  • Open source — MIT license, same as OpenClaw
  • 30-second migration — If you use OpenClaw, npm install -g soulclaw and you're done

SoulClaw is a fork of OpenClaw that adds memory and safety features. Your existing ~/.openclaw/ configuration, SOUL.md, workspace, and channel tokens all transfer automatically.

Migration 

npm install -g soulclaw
soulclaw gateway install
soulclaw gateway start  # Uses existing ~/.openclaw/ config
Enter fullscreen mode Exit fullscreen mode

Full guide: Migration from OpenClaw

What's Next

We're building towards a vision where AI agents are:

  1. Safe — You know what they're doing and they stay in character
  2. Remembered — Nothing is lost, ever (3-Tier Memory + DAG)
  3. Collaborative — Agents share knowledge across devices

The CLI guide has full documentation for all commands.

SoulClaw v2026.3.21 — Safe agents with long-term memory.
GitHub · npm · Docs

Originally published at https://blog.clawsouls.ai/posts/soulclaw-safe-agents/

Top comments (0)