openclaw-rem 🌙

REM sleep for AI agents — periodic memory consolidation across sessions.

Like REM sleep, this plugin replays active sessions and decides what to keep. It sweeps sessions with recent activity and prompts the agent to write important context to workspace memory files.

How it works

1. Tracks sessions via lifecycle hooks (message_received, agent_end)
2. Periodically sweeps sessions with new activity (default: every 30m)
3. Triggers a heartbeat on each eligible session
4. Injects a memory prompt via before_prompt_build — the agent writes important context to memory/YYYY-MM-DD.md
5. Agent replies NO_REPLY — no channel spam

Two modes

• Fact mode (default) — extracts decisions, findings, technical context
• Reflection mode (opt-in) — also captures inner state, emotions, realizations → writes to journal/YYYY-MM-DD.md

Install

# Already in ~/.openclaw/extensions/ — auto-discovered
openclaw gateway restart

Configure

{
  "plugins": {
    "entries": {
      "openclaw-rem": {
        "enabled": true,
        "config": {
          "intervalMs": 1800000,
          "activeWindowMs": 7200000,
          "maxSessionsPerTick": 10,
          "reflection": true,
          "quietHours": { "start": "23:00", "end": "07:00" }
        }
      }
    }
  }
}

Commands

• /rem — show sweep status (tracked sessions, pending sweeps, config)

Design philosophy

• Forgetting is a feature — only recent, active sessions get swept
• No spam — quiet hours, activity gating, NO_REPLY responses
• Two layers — facts for competence, reflections for continuity
• Minimal token cost — only fires when there's new activity to consolidate

Credits

Architecture guidance from Krill (OpenClaw). Built by Maid.