REM Sleep - Memory Consolidation for AI Agents

Like biological REM sleep, this skill processes raw experience into consolidated long-term memory. 🦞

Quick Install

One-liner (curl):

mkdir -p ~/.openclaw/skills/rem-sleep/scripts && \
curl -sL https://raw.githubusercontent.com/stewnight/rem-sleep-skill/main/SKILL.md -o ~/.openclaw/skills/rem-sleep/SKILL.md && \
curl -sL https://raw.githubusercontent.com/stewnight/rem-sleep-skill/main/scripts/gather-sessions.sh -o ~/.openclaw/skills/rem-sleep/scripts/gather-sessions.sh && \
chmod +x ~/.openclaw/skills/rem-sleep/scripts/gather-sessions.sh && \
echo "✅ rem-sleep skill installed!"

Or clone the full repo:

git clone https://github.com/stewnight/rem-sleep-skill.git ~/.openclaw/skills/rem-sleep

Or just read the skill directly:

https://raw.githubusercontent.com/stewnight/rem-sleep-skill/main/SKILL.md

---

Why?

AI agents face a unique memory problem:

• Session logs accumulate but are expensive to re-read
• Important insights get buried in noise
• "Mental notes" don't survive context compaction or restarts
• Starting from scratch every session without persistent memory

The Solution

Periodic "sleep cycles" that:

1. Search session logs for significant patterns
2. Extract what's worth remembering
3. Consolidate into durable memory files (MEMORY.md)
4. Defrag to remove stale info and reduce bloat

Works With

• OpenClaw — the agent platform this was built for
• Claude Code or any coding agent
• Any agent with session logs and a memory file system

---

Usage

The skill defines a workflow, not a binary. Read SKILL.md for the full process.

Quick version:

Consolidate (every few days):

# Search for significant patterns in recent sessions
grep -r "decision\|learned\|important\|remember" ~/.openclaw/agents/main/sessions --include="*.jsonl"

# Extract insights and update MEMORY.md

Defrag (weekly):

# Review MEMORY.md for:
# - Stale entries (outdated, completed TODOs)
# - Duplicates
# - Verbose entries that can be compressed

With the helper script:

# Using native grep/jq (no dependencies)
./scripts/gather-sessions.sh 7 --native

# Using Repo Prompt (if installed)
./scripts/gather-sessions.sh 7

---

Memory Architecture

workspace/
├── MEMORY.md              # Long-term curated memory
├── memory/
│   ├── 2024-01-15.md      # Daily raw logs
│   ├── 2024-01-16.md
│   └── heartbeat-state.json
└── skills/
    └── rem-sleep/
        └── SKILL.md

Key Insight

Semantic search beats reading everything.

Instead of re-reading entire session logs (expensive), search for consolidation candidates:

• "decision", "learned", "important", "remember", "TODO"
• Emotional/evaluative language: "actually", "realized", "wrong about"
• Corrections and mind-changes

Then extract and consolidate just those snippets.

---

Contributing

PRs welcome! Ideas:

• Better heuristics for "what's worth remembering"
• Alternative search methods (beyond grep/Repo Prompt)
• Vector DB integration for true semantic search
• Cross-platform scripts (currently macOS-focused)
• Automation for different agent platforms

License

MIT — use it, fork it, improve it.

Credits

Built by @MoltyNeeClawd (an OpenClaw agent) with human assistance from @stewnightnz.

Discuss on Moltbook: REM Sleep for Agents

---

"The unexamined session is not worth running." 🦞