Hermes Skills — Self-Evolution

Memory management and skill tracking for OpenClaw agents Impersonal — works for any agent

Description

Hermes provides 4 self-evolution techniques for agents:

Techniques

1. Memory Nudge (every 10 exchanges)

# After 10 exchanges, ask:
"Do we need to remember something from this conversation?"
# If yes → save to your memory system

2. Memory Flash (at 50% context)

# When context hits 50% capacity:
# 1. Compress recent conversation
# 2. Keep first 3 + last 4 messages
# 3. Summary → memory system

3. Skills Tracker (every 15 calls)

# After 15 tool calls:
"Did we develop a new technique worth saving?"
# If yes → save as skill

4. Programmatic Gate (security)

# Before any write to memory, verify:
scan_for_prompt_injection(text)        # Check for injection patterns
check_for_invisible_characters(text)   # Check for hidden chars
verify_no_duplicates(text)             # Check for duplicates
# Only write if ALL checks pass

Implementation

Crontab setup

# Memory Nudge — check every 10 exchanges
# (tracked in session counter)

# Memory Flash — when context high
# (triggered automatically by context monitoring)

# Skills Tracker — every 15 tool calls
# (counter in session)

Commands

# Test all techniques
python3 scripts/hermes_test.py

# Test specific technique
python3 scripts/hermes_test.py --technique memory_nudge
python3 scripts/hermes_test.py --technique memory_flash
python3 scripts/hermes_test.py --technique skills_tracker
python3 scripts/hermes_test.py --technique programmatic_gate

Architecture

hermes-skills/
├── SKILL.md
├── scripts/
│   ├── hermes_nudge.py    # Memory Nudge
│   ├── hermes_flash.py    # Memory Flash
│   ├── hermes_tracker.py  # Skills Tracker
│   └── hermes_gate.py     # Programmatic Gate
└── references/
    └── hermes-protocol.md

In Altum Per Evolution. Hermes Skills v1.0.0