hermes-memory-keep-alive

name	hermes-memory-keep-alive
description	Automatic task memory and keep-alive loop for Hermes agents. Every task gets durable RESUME / CHECKLIST / DOCS notes in Obsidian. For long tasks, arm the keep-alive loop to detect stalls and keep work moving — then disarm when done so you're not burning tokens.
metadata	{"author":"terry","version":"1.1.0","hermes":{"tags":["workflow","memory","recovery","restart-safe","obsidian","loop","keep-alive"],"related_skills":[]}}

Hermes Memory + Keep-Alive

Two things in one skill:

1. Task Memory — every task automatically gets persistent notes so work survives restarts
2. Keep-Alive Loop — arm it for long tasks, disarm when done

Task Memory

When you receive any task, automatically create a task folder in the Obsidian vault:

<vault>/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.

Vault path:

Use OBSIDIAN_VAULT_PATH if set, otherwise default to ~/Documents/Obsidian Vault. Always quote vault paths (they may contain spaces).

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 Tasks/Session-Resume-Workflow/LOOP-STATE.md.

Monitoring layers

When armed, three jobs watch your agent's work:

Layer	Job name	Schedule	What it does
Watchdog	keep-alive-watchdog	15m	Detects stalls, writes WATCHDOG.md
Replayer	keep-alive-replayer	30m	Takes one concrete step on a stalled task
Escalator	keep-alive-escalator	60m	Forces fresh-session handoff on repeated stalls

Integrity layers

These run always, regardless of loop state:

Layer	Job name	Schedule	What it does
Validator	memory-validator	60m	Repairs missing notes, refreshes workflow index
Smoke test	memory-smoke-test	360m	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. Do not invent formats.
• Refresh the workflow 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 workflow notes exist (TEMPLATE.md, WORKFLOW-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)
• Hermes memory provider is healthy (runs hermes memory status, Hermes 0.7.0+)

Validator checks:

• Repairs missing RESUME.md, CHECKLIST.md, DOCS.md from canonical template
• Refreshes WORKFLOW-INDEX.md with current task state
• Checks Hermes built-in MEMORY.md character count against the 2,200-char limit — warns at 80% capacity (Hermes 0.7.0+)

File layout

<vault>/Tasks/
  Session-Resume-Workflow/
    TEMPLATE.md          # canonical template for new task folders
    WORKFLOW-INDEX.md    # quick-scan index of all tasks
    LOOP-STATE.md        # armed/disarmed state marker
  <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.