Memory Boost
Two things in one skill:
- Task Memory — every task automatically gets persistent notes so work survives restarts
- Keep-Alive Loop — arm it for long tasks, disarm when done
No external dependencies. Everything lives inside ~/.openclaw/memory/.
Task Memory
When you receive any task, automatically create a task folder:
~/.openclaw/memory/tasks/<task-name>/
RESUME.md — status, heartbeat, next action, key files, restart note
CHECKLIST.md — step-by-step progress
DOCS.md — goal, decisions, gotchas, notes for next session
RESUME.md must contain:
- Task name and one-line description
- Current status (active / stalled / complete)
- Last heartbeat (date and time, updated whenever you touch the task)
- Next action (one concrete step)
- Key files (relevant paths)
- Restart note (what a fresh session needs to pick this up)
CHECKLIST.md must contain:
- Numbered or checkbox steps
- Current status of each step
- A final verification step
DOCS.md must contain:
- Goal
- Important decisions and why
- File paths that matter
- Gotchas and failure modes
- Notes for the next session
Update rules:
- Create notes before starting a multi-step task.
- Update RESUME.md before and after major milestones.
- Mark checklist items complete as soon as they are done.
- Add discoveries and gotchas to DOCS.md immediately.
- Always update the heartbeat when touching an active task.
- A heartbeat older than 24 hours or a missing next action means the task is stale.
Storage path:
Task notes live in ~/.openclaw/memory/tasks/. The canonical template is at {baseDir}/templates/TEMPLATE.md.
Keep-Alive Loop
For long-running tasks, the keep-alive loop prevents silent failures.
Controls
/loop-start — arms the loop. Monitoring jobs begin running./loop-stop — disarms the loop. Jobs no-op. No wasted tokens.
The loop state lives in ~/.openclaw/memory/LOOP-STATE.md.
Monitoring layers
When armed, three jobs watch your agent's work:
| Layer | Job name | Schedule | What it does |
|---|
| Watchdog | boost-watchdog | */15 * * * * | Detects stalls, writes WATCHDOG.md |
| Replayer | boost-replayer | */30 * * * * | Takes one concrete step on a stalled task |
| Escalator | boost-escalator | 0 * * * * | Forces fresh-session handoff on repeated stalls |
Integrity layers
These run always, regardless of loop state:
| Layer | Job name | Schedule | What it does |
|---|
| Validator | boost-validator | 5 * * * * | Repairs missing notes, refreshes task index |
| Smoke test | boost-smoke-test | 0 */6 * * * | Verifies the skill itself is healthy |
Watchdog rules:
- Write WATCHDOG.md only when you can prove a stall. Do not manufacture problems.
- Include: folder path, blocker, why-stalled tag, one next action.
- Why-stalled tags:
blocked-on-external, ambiguous-next-step, repeated-promise, missing-context
Replayer rules:
- Take only one mechanical step per pass. No speculative fixes.
- If the step is ambiguous, update WATCHDOG.md with a clearer next action and stop.
Escalator rules:
- Escalate only after the same stall repeats or a replayer pass fails to advance.
- Add an ESCALATE section to WATCHDOG.md with what was tried and the strongest next action.
Validator rules:
- Inspect task folders for missing RESUME.md, CHECKLIST.md, or DOCS.md.
- Backfill missing notes from the canonical template at
{baseDir}/templates/TEMPLATE.md. Do not invent formats.
- Refresh the task index so it lists all tasks with status and next action.
- Treat a task as stale if heartbeat > 24h or next action is missing.
Smoke test checks:
- Core files exist (TEMPLATE.md, TASK-INDEX.md, LOOP-STATE.md)
- All 5 scheduled jobs exist and are enabled
- At least one active task has a heartbeat and next action
- Loop state is valid (armed or disarmed)
File layout
~/.openclaw/memory/
TEMPLATE.md # canonical template for new task folders
TASK-INDEX.md # quick-scan index of all tasks
LOOP-STATE.md # armed/disarmed state marker
tasks/
<task-name>/
RESUME.md
CHECKLIST.md
DOCS.md
WATCHDOG.md # created by watchdog when stall detected
When to use
- Task memory: Always. Every task should get notes automatically.
- Keep-alive loop: For long tasks.
/loop-start before the task, /loop-stop when done.
- Quick tasks: Just let the task memory handle it. No need to arm the loop.
