openclaw-dream-setup

name	openclaw-dream-setup
description	Set up nightly dream routine (memory distillation) for an OpenClaw agent including DREAM-ROUTINE.md, cron job, and archive pipeline
argument-hint	["agent-id"]
disable-model-invocation	true

Set Up Dream Routine

Configure nightly memory distillation for agent $ARGUMENTS.

Reference

Read ~/.openclaw/workspace/skills/02-memory-system.md for the full memory system guide.

Setup Detection

OPENCLAW_REPO=$(readlink ~/.openclaw/openclaw.json 2>/dev/null | sed 's|/.openclaw/openclaw.json||')
TIER_CONFIGS=(~/.openclaw/configs/openclaw-*.json)
[[ -f "${TIER_CONFIGS[0]}" ]] && MULTI_GATEWAY=true || MULTI_GATEWAY=false

If multi-gateway, resolve the agent's tier and config:

for cfg in ~/.openclaw/configs/openclaw-*.json; do
  if grep -q "\"id\": *\"$ARGUMENTS\"" "$cfg" 2>/dev/null; then
    TIER=$(basename "$cfg" | sed 's/openclaw-//;s/.json//')
    TIER_CONFIG="$OPENCLAW_REPO/.openclaw/configs/openclaw-$TIER.json"
    break
  fi
done

Steps

1. Create directories:

mkdir -p "$OPENCLAW_REPO/.openclaw/agents/$ARGUMENTS/memory/archives"

2. Ask the user:

   • What time should the dream routine run? (default: 23:00)
   • What timezone? (default: user's local timezone)
   • Any specific focus areas for distillation? (e.g., "focus on project status and commitments")

3. Create DREAM-ROUTINE.md at $OPENCLAW_REPO/.openclaw/agents/$ARGUMENTS/DREAM-ROUTINE.md:

# DREAM-ROUTINE.md

## Trigger
Nightly cron at <time> <timezone>.

## Process
1. Read today's `memory/YYYY-MM-DD.md`
2. Read `MEMORY.md` for long-term context
3. Distill to `memory/YYYY-MM-DD-DISTILLED.md` (max 2,500 tokens)
4. Update `memory/MEMORY-DIGEST.md` (3-day rolling window, max 7,500 tokens)
5. Archive distillations older than 3 days to `memory/archives/`

## Distillation Format
# Distilled — YYYY-MM-DD

## Decisions
- <decision, who, context>

## Project Updates
- <project>: <change>

## New Context
- <new info worth remembering>

## Completed
- <finished tasks>

## Blockers
- <unresolved issues>

## Tomorrow
- <items needing attention>

## Rules
- NEVER include credentials or secrets
- Stay within 2,500 token budget
- Focus on what CHANGED, not status quo
- If no daily log exists, skip gracefully

4. Create MEMORY.md if it doesn't exist — a starter template at $OPENCLAW_REPO/.openclaw/agents/$ARGUMENTS/MEMORY.md:

# MEMORY.md — <Agent Name>

## Active Projects
(To be filled as the agent works)

## Key Contacts
(To be filled)

## Standing Rules
(To be filled)

5. Add cron job to $OPENCLAW_REPO/.openclaw/cron/jobs.json. Generate a UUID first:

python3 -c "import uuid; print(uuid.uuid4())"

Add the job:

{
  "id": "<generated-uuid>",
  "agentId": "$ARGUMENTS",
  "name": "<Agent Name> Dream Routine",
  "enabled": true,
  "schedule": {
    "kind": "cron",
    "expr": "0 <hour> * * *",
    "tz": "<timezone>"
  },
  "sessionTarget": "isolated",
  "wakeMode": "now",
  "payload": {
    "kind": "agentTurn",
    "message": "Run your nightly dream routine. Read DREAM-ROUTINE.md and follow the process exactly. Read today's daily log from memory/, distill it, update the rolling digest, and archive old distillations.",
    "model": "anthropic/claude-sonnet-4-5",
    "timeoutSeconds": 120
  },
  "delivery": {
    "mode": "announce",
    "channel": "last"
  }
}

6. Add QMD paths to the appropriate config:

• Multi-gateway: edit $TIER_CONFIG
• Single-gateway: edit $OPENCLAW_REPO/.openclaw/openclaw.json

Add to memory.qmd.paths:

{
  "path": "$HOME/.openclaw/agents/$ARGUMENTS/memory",
  "name": "$ARGUMENTS-memory",
  "pattern": "**/*.md"
},
{
  "path": "$HOME/.openclaw/agents/$ARGUMENTS",
  "name": "$ARGUMENTS-docs",
  "pattern": "*.md"
}

Replace $HOME with the user's actual home directory path.

7. Update AGENTS.md — Add memory loading to the session startup sequence in $OPENCLAW_REPO/.openclaw/agents/$ARGUMENTS/AGENTS.md:

## Every Session
1. Read `SOUL.md`
2. Read `USER.md`
3. Read `MEMORY.md` — curated long-term context
4. Read `memory/MEMORY-DIGEST.md` — rolling 3-day digest (if exists)
5. Do NOT load raw daily logs on startup

8. Stow:

rm -f ~/.openclaw/cron/jobs.json
cd "$OPENCLAW_REPO" && stow --no-folding -t ~ .

9. Verify: The dream routine will run at the next scheduled time. Check logs:

• Multi-gateway: tail -30 ~/.openclaw-$TIER/gateway.log
• Single-gateway: tail -30 ~/.openclaw/logs/gateway.log

Important

• Token budgets are critical: 2,500/day distillation, 7,500 rolling digest
• Daily logs (memory/YYYY-MM-DD.md) should be gitignored — they're ephemeral
• MEMORY.md is curated and committed — it's the permanent knowledge base
• Stagger dream times across agents (e.g., 20:30, 20:35, 20:40) to avoid concurrent load