lucid-dreamer

Nightly AI memory reasoning system. Lucid runs every night while you sleep - it reads your daily notes and memory files, detects stale facts, unresolved todos, recurring problems, forgotten decisions, and can optionally perform aggressive cleanup and contradiction detection. Includes optional session debrief for quick end-of-day memory capture. Zero dependencies, no database, no embeddings. Just a scheduled job and markdown files. Use when you want your AI agent to automatically maintain and improve its long-term memory over time. Triggers on "memory dreamer", "nightly memory review", "lucid", "auto memory", "memory cleanup", "memory hygiene".

Your AI sleeps. Lucid dreams.

Lucid keeps your AI's memory clean. Every night, it reads what happened, checks what your AI already knows, and suggests what's outdated, missing, or forgotten.

See README.md for full setup, ARCHITECTURE.md for internals, and config/ for configuration.

OpenClaw 2.0: check the built-in memory sweep first

OpenClaw 2.0 ships Grounded Dreaming in the memory-core plugin and enables it by default. A fresh install already has this automation:

text

memory-core:memory-dreaming   cron 0 3 * * * (exact)   isolated

It writes MEMORY.md on the same schedule Lucid has historically used. Two unattended processes rewriting the same file every night will lose data.

Pick one of these before you schedule anything:

• Run Lucid instead. Turn the built-in sweep off by running /dreaming off in a chat session, then schedule Lucid at 03:00 as before.
• Run both, separated. Keep the built-in sweep and schedule Lucid at a different hour (04:30 works). Lucid still writes its own report to memory/review/, so the two do not collide as long as Lucid's auto-apply stays off - which is the default.
• Run only the built-in sweep. Lucid's review reports and approval workflow are the reason to keep it; the built-in sweep has no review gate.

Check what is scheduled on your install with openclaw automations list.

Quick Setup

1. Set your workspace path in the config:

   bash

   export CLAWD_DIR=/path/to/your/workspace
   

2. Schedule the nightly job with openclaw automations add, running the prompt in prompts/nightly-review.md. Read the section above before choosing the hour.

   Optional: add a lightweight session debrief automation around 18:00 using prompts/session-debrief.md. This is a faster daily capture pass than the nightly review — it reads today's daily note and writes durable decisions/facts straight into memory without creating a review report.

3. Wake up to a review report in memory/review/YYYY-MM-DD.md.

4. Approve or reject suggestions — Lucid tracks state in memory/review/state.json.

Optional Session Debrief Cron

Use prompts/session-debrief.md for a quick end-of-day memory pass around 18:00. It is designed to run faster than the nightly review: read today's daily note, capture durable decisions/facts/action items, and write them directly into memory.

Recommended settings. openclaw automations is the current command; openclaw cron still works as an alias. Use a fully qualified provider/model reference - short aliases are skipped without an error and the job falls back to the session default.

bash

openclaw automations add \
  --name "lucid-debrief" \
  --cron "0 18 * * *" \
  --tz "Europe/Vienna" \
  --model "anthropic/claude-haiku-4-5" \
  --session isolated \
  --message "$(cat prompts/session-debrief.md)"

What it does:

• Reads today's daily note (memory/TODAY.md)
• Captures key decisions, durable facts, and concrete action items
• Writes those updates directly into long-term memory
• Skips the full review report to stay quick and cheap

Files

• prompts/nightly-review.md — the main nightly review prompt
• prompts/session-debrief.md — optional quick-capture prompt for ~18:00
• config/ — thresholds and behavior settings
• examples/ — sample review output and state file

Security

Files read at runtime:

• MEMORY.md — long-term agent memory summary
• USER.md — user profile and preferences
• Last 7 daily notes (memory/YYYY-MM-DD.md)

Files written at runtime:

• memory/review/YYYY-MM-DD.md — the generated review report
• memory/review/state.json — approval/rejection tracking state

What this skill is designed to avoid:

• Avoid suggesting or outputting passwords, API keys, tokens, or other credentials in generated memory updates
• Never accesses files outside the configured workspace directory
• Never pushes to remote git automatically — all commits are local only, and no git push is performed unless you explicitly run it
• Announce/notification delivery is opt-in and off by default — no messages are sent without explicit configuration

Recommendations:

• Set CLAWD_DIR explicitly in your environment to ensure the skill operates on the correct workspace
• This skill reads workspace markdown files such as MEMORY.md, USER.md, and recent daily notes. Do not run it on a workspace containing unencrypted API keys or other secrets in plain markdown files.
• Review generated reports before approving suggestions — Lucid proposes changes, but you remain in control