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 todo...

Security Scan

Lucid Dreamer appears to do what it says (nightly analysis of markdown-based workspace memory, optional local git commits, trend detection via included Python script); the requested binaries and behavior align with its purpose, with a few documentation/metadata inconsistencies you should confirm before enabling auto-apply/cleanup on a real workspace.

Purpose & Capability

Name/description match the implementation: it reads workspace markdown, produces a review, optionally edits MEMORY.md and commits locally. Required binaries (git, date, python3) are appropriate for the stated behaviors (local git commits, date calculation, trend script). Minor documentation mismatch: SKILL.md and runtime checks require CLAWD_DIR to be set, but registry metadata lists no required env vars — this is a documentation/metadata inconsistency (not necessarily malicious).

Instruction Scope

SKILL.md instructions are narrowly scoped to a workspace directory: read USER.md, MEMORY.md/sections, last 7 daily notes; do not read memory/review/*.md; write review files and state.json; run an included trend_detection.py. It explicitly forbids reading credentials and forbids auto-pushing to remotes. The instructions contain git commit steps and file edits, which are coherent with the stated auto-apply/cleanup features.

Install Mechanism

Instruction-only skill with included helper scripts (no install spec). No network downloads or extract/install steps are present in the registry metadata. This lowers install risk; included Python scripts will be present in the workspace but not automatically executed outside the described runtime steps.

Credentials

The skill does not request external credentials or secrets and operates on local workspace files only, which is proportionate. However, it relies on CLAWD_DIR being set and will abort if unset — yet CLAWD_DIR is not declared as a required env var in registry metadata. Also review notifications/announce settings (they default off) and ensure you don't configure a delivery channel you don't trust. The skill warns not to keep secrets in plain markdown — follow that guidance.

Persistence & Privilege

No elevated or persistent system privileges requested. always: false. The skill writes only to workspace paths (review files, state.json, .last-success) and performs local git commits; it does not auto-push to remotes by default. Auto-apply and aggressive cleanup are opt-in in config, so the default behavior is conservative.

Assessment

What to check before installing/using Lucid Dreamer: - Confirm CLAWD_DIR handling: SKILL.md aborts if CLAWD_DIR is unset, but the registry metadata doesn't list CLAWD_DIR as a required env var — set CLAWD_DIR explicitly and test the skill in a sandbox workspace first. - Test in a disposable workspace: run the nightly prompt against a copy of your memory files to see exact edits and commits before enabling on real data. - Inspect included scripts (scripts/trend_detection.py and scripts/migrate_memory.py) yourself or run them in a controlled environment to ensure no unexpected network I/O; the package claims stdlib-only, but verify. - Review config files: autoApply.enabled and aggressiveCleanup.enabled default to false; leave them off until you're confident. If you enable auto-apply or aggressive cleanup, review config/auto-apply.md and config/lucid.config.json to limit categories. - Check git remotes and hooks: although the skill promises not to push, local commits will be created. Make sure your workspace git has no unexpected remotes or hooks that could cause remote pushes or side effects. - Keep secrets out of workspace markdown: the skill explicitly reads USER.md and MEMORY.md; do not store API keys/passwords/tokens in plaintext in those files. - Consider limiting autonomous runs: the skill can be scheduled via cron; if you prefer manual control, avoid adding the cron job or disable autonomous invocation until you have validated behavior. Overall: behavior is coherent with its purpose; the main risks are accidental, not malicious — use careful testing, explicit CLAWD_DIR, and conservative config before enabling auto-apply/cleanup on production memories.

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.

Quick Setup

1. Set your workspace path in the config:

   export CLAWD_DIR=/path/to/your/workspace
   

2. Create a nightly cron job using OpenClaw's cron tool — run the prompt in prompts/nightly-review.md at 3 AM.

   Optional: add a lightweight session debrief cron 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 OpenClaw cron settings:

openclaw cron add \
  --name "lucid-debrief" \
  --cron "0 18 * * *" \
  --tz "Europe/Vienna" \
  --model "your-preferred-model" \  # e.g. anthropic/claude-haiku-4-5 or opencode-go/minimax-m2.7
  --session isolated \
  --wake-mode now \
  --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