🧠 Squad Memory

Persistent memory for OpenClaw sub-agent squads.

Sub-agents are stateless — they wake up fresh every session with zero recall. Squad Memory gives them three layers of persistent knowledge:

Layer	What	Lifespan	Flushable?
Process	How you work, standards, values	Permanent	Never
Program	Project expertise, role patterns	Per-project	Optional
Episodic	Session logs, recent context	Rolling	Yes

Why?

Without persistent memory, your orchestrator spends thousands of tokens re-explaining context every spawn. Sub-agents repeat mistakes. They can't learn from experience.

With Squad Memory:

• 50% fewer context tokens per spawn
• Role-tagged learnings compound across sessions
• Task-aware selection loads only relevant memories
• Flush controls let you start fresh without losing expertise
• Dynamic role discovery adapts to your squad structure

Install

# Clone into your OpenClaw skills directory
git clone https://github.com/clawbunny/squad-memory.git ~/.openclaw/skills/squad-memory

# Initialize process memory
mkdir -p ~/.openclaw/workspace/memory/squads/_process
cp ~/.openclaw/skills/squad-memory/process-template.md ~/.openclaw/workspace/memory/squads/_process/standards.md

Usage

SM="~/.openclaw/skills/squad-memory/squad-memory.sh"

# Write after a spawn completes
echo "## Session summary..." | $SM write my-squad -

# Read before spawning (inject into task prompt)
$SM read my-squad --tokens 500

# Smart selection for architecture task
$SM read my-squad --task "architecture" --role architect --tokens 400

# Distill patterns from session history
$SM distill my-squad

# Start fresh (archives first)
$SM flush my-squad --keep-semantic

How It Works

Before Spawn:
  read → Process + Program + Episodic → inject into task prompt (500 tokens)

After Spawn:
  Extract "Session Memory" → write → role-tagged learnings accumulate

Every 3 Sessions:
  Auto-distill → episodic patterns promoted to semantic/program memory

Example Role Tags

Use any role names that fit your squad. The system dynamically discovers them:

## Session Memory
- [ARCHITECT] Learned to validate schema before migration
- [ANALYST] Market research shows 60% prefer feature X
- [COORDINATOR] Sprint planning works better async for distributed teams
- [ALL] Always test edge cases before deployment

Security

Version 2.0 includes comprehensive security hardening:

• ✅ Path traversal protection (squad IDs validated)
• ✅ Command injection prevention (all expansions quoted)
• ✅ Race condition mitigation (mktemp for temp files)
• ✅ Input validation (numeric parameters verified)
• ✅ Safe glob handling (no unsafe wildcards)
• ✅ ShellCheck compliant

Tested

Test	Result
Basic write-read cycle	✅
Squad isolation (no cross-contamination)	✅
Token budget enforcement	✅
Live recall (5/5 questions from memory)	✅
Task-aware selection	✅
Semantic distillation (5 sessions → patterns)	✅
Flush validation (6/6 with zero episodic)	✅
Cross-session knowledge transfer	✅
Path traversal prevention	✅
Dynamic role discovery	✅

License

MIT

Links

• GitHub: https://github.com/clawbunny/squad-memory
• ClawHub: Coming soon