Agent Ops Toolkit Skill

Overview

Agent Ops Toolkit sets up the operational backbone for any OpenClaw agent. Five core components work together to keep your agent learning, accountable, and in sync with your schedule:

1. Nightly Extraction — Consolidates conversations and decisions into atomic facts
2. Morning Brief — Daily priorities and overnight activity summary
3. Heartbeat Monitoring — Health checks for managed agent loops (stall detection, auto-restart)
4. Memory Decay — Kalman-inspired lifecycle management (hot → warm → cold)
5. PARA Scaffold — Ready-to-use knowledge graph structure

Time to setup: 5 minutes (wizard-driven)
Ongoing maintenance: ~1 minute/day (morning brief review)
Cost: ~$5–10/month on Claude Haiku extraction models

All components are research-backed and cost-optimized for 24/7 autonomous operation.

Setup Wizard

The fastest way to get started:

clawhub install ops-toolkit
bash scripts/setup_ops.sh

The interactive wizard prompts for:

1. Timezone (default: America/New_York)

   • Used for scheduling nightly extractions and morning briefs

2. Telegram Chat ID (optional)

   • Where morning briefs are delivered
   • Leave blank to skip morning brief setup

3. Agent ID (default from OpenClaw config)

   • Identifies your agent in logs and cron jobs

4. Extraction Model (default: anthropic/claude-haiku-4-5)

   • Fast, cheap model for nightly fact extraction
   • Haiku handles structured tasks at 1/60th the cost of Opus

5. Morning Brief Model (default: anthropic/claude-haiku-4-5)

   • Model for synthesis and summary generation
   • Haiku is sufficient for brief writing; upgrade to Sonnet if you want richer prose

Output: Two ready-to-use cron configs (nightly-extraction-cron.json, morning-brief-cron.json) and your PARA scaffold directories.

Next steps shown on-screen; you manually run the openclaw cron add commands (for safety, not auto-executed).

Component Architecture

1. Nightly Extraction (scripts/heartbeat_tick.py)

What it does:

• Runs on schedule (default: 11 PM in your timezone)
• Reads conversation history from your daily notes
• Extracts key facts, decisions, goals, and learnings
• Writes atomic facts to life/items.json with metadata

Why it matters:

• Consolidates raw experience into retrievable facts
• Removes burden of manual logging
• Enables decay algorithm to age memories appropriately

Configuration: templates/nightly-extraction-cron.json

See also: references/memory-schema.md for items.json spec

2. Morning Brief (scripts/decay_sweep.py)

What it does:

• Runs on schedule (default: 8 AM in your timezone)
• Generates curated summary from hot/warm facts and goals
• Delivers to your Telegram chat (or stdout if unconfigured)
• Shows priorities, overnight activity, risks

Why it matters:

• Saves 5 minutes of manual context gathering each morning
• Keeps agent aligned with your goals
• Surfaces newly-cold or risky facts

Configuration: templates/morning-brief-cron.json

See also: references/decay-algorithm.md for hot/warm/cold definitions

3. Heartbeat Monitoring (scripts/heartbeat_tick.py)

What it does:

• Runs every 30 minutes (configurable)
• Checks managed tmux sessions for progress or stalls
• Detects when output hasn't changed (stall detection via hash)
• Outputs HEARTBEAT_OK if healthy, or ALERT: <message> + NEXT: <action> if intervention needed

Why it matters:

• Autonomous loops can hang without visibility
• Early stall detection prevents silent failures
• Auto-restart capability for managed sessions

Configuration: templates/heartbeat-config.json

See also: references/heartbeat-protocol.md for protocol details

4. Memory Decay (scripts/decay_sweep.py)

What it does:

• Runs weekly (default: Sunday at 2 AM)
• Classifies facts as hot/warm/cold based on lastAccessed field
• Hot facts (accessed < 7 days) remain prominent in summaries
• Warm facts (8–30 days) lower in priority
• Cold facts (> 30 days) removed from summaries but kept in storage
• Frequently-accessed facts (accessCount > 5) get 14-day resistance bonus

Why it matters:

• Without decay, memory becomes a graveyard of irrelevant facts
• Decay surfaces active concerns while preserving historical record
• Access-frequency resistance means "living" facts don't age

Algorithm: Inspired by GAM-RAG Kalman principle: "fast warm-up for novel signals, conservative refinement for stable ones"

See also: references/decay-algorithm.md for formal rules and cite

5. PARA Scaffold

What it does:

• Creates ready-to-use directory structure for knowledge graph:
  • projects/ — active initiatives
  • areas/ — ongoing responsibilities (people, companies, expertise domains)
  • resources/ — reference material (papers, tools, templates)
  • archives/ — completed/inactive items

Why it matters:

• PARA is battle-tested for long-term personal knowledge management
• Pre-built structure removes decision paralysis
• Atomic facts (items.json) naturally organize into PARA entities

See also: Your Agent's Memory chapter in the quickstart guide

Generation Flow

After setup:

1. Wizard generates nightly-extraction-cron.json with your timezone, agent ID, model choice
2. Wizard generates morning-brief-cron.json with your Telegram chat ID (optional)
3. Wizard creates PARA scaffold: life/{projects,areas,resources,archives}/
4. You run: openclaw cron add < nightly-extraction-cron.json
5. You run: openclaw cron add < morning-brief-cron.json
6. Cron jobs activate on next scheduled tick
7. Nightly: facts extracted, items.json updated, summaries rewritten
8. Morning: brief composed and delivered
9. Weekly: decay sweep ages cold facts

All configs are human-editable. Change models, timezones, or delivery channels anytime.

Model Routing (Cost Optimization)

The toolkit uses cost-conscious model selection informed by MemPO (arXiv:2603.00680):

Result: Month of nightly extraction + morning briefs ≈ $5–10 vs $300+ with Opus.

Self-managed memory (MemPO) reduces token usage 67–73%, making this feasible.

Research Context

All design choices are informed by peer-reviewed research:

GAM-RAG (arXiv:2603.01783)

Finding: Kalman-inspired updates apply rapid changes to uncertain memories, conservative refinement to stable ones.
Application: Decay algorithm — new facts (uncertain) update easily; established facts (stable) resist change via access-count resistance.

SuperLocalMemory (arXiv:2603.02240)

Finding: Local-first, 4-layer progressive architecture with Bayesian trust scoring and provenance tracking.
Application: items.json schema includes source, timestamp, accessCount (trust signals); stored locally, never cloud-synced.

Retrieval Bottleneck (arXiv:2603.02473)

Finding: Retrieval quality (which facts you surface) matters 20× more than write sophistication (how fancy your summaries are).
Application: Store raw atomic facts, rely on vector search and decay ranking. Skip expensive summarization.

MemPO (arXiv:2603.00680)

Finding: Self-managed memory reduces token cost 67–73% without sacrificing quality.
Application: Agent autonomously prunes/prioritizes via decay; uses cheap extraction models (Haiku); no expensive fine-tuning.

See references/research-notes.md for full citations and deeper design mappings.

Templates & Scripts

Templates

• nightly-extraction-cron.json — Parameterized with {{TIMEZONE}}, {{AGENT_ID}}, {{MODEL}}
• morning-brief-cron.json — Parameterized with {{TIMEZONE}}, {{AGENT_ID}}, {{MODEL}}, {{DELIVERY_CHANNEL}}, {{CHAT_ID}}
• heartbeat-config.json — Default heartbeat configuration
• life-scaffold/ — PARA directory structure

Scripts

• setup_ops.sh — Interactive wizard (bash)
• heartbeat_tick.py — Stall detection + restart logic
• decay_sweep.py — Weekly fact lifecycle processor

References

• memory-schema.md — Full items.json specification
• cron-templates.md — Documented cron configs
• heartbeat-protocol.md — Deterministic heartbeat protocol
• decay-algorithm.md — Formal decay rules with formula
• research-notes.md — Paper citations and mappings

Quick Start

# 1. Install the skill
clawhub install ops-toolkit

# 2. Run the setup wizard
bash scripts/setup_ops.sh

# 3. Follow the prompts (timezone, Telegram ID optional, model choice)

# 4. The wizard outputs next steps, e.g.:
# "To activate, run:"
#   openclaw cron add < nightly-extraction-cron.json
#   openclaw cron add < morning-brief-cron.json

# 5. Run those commands manually (wizard doesn't auto-execute for safety)

# 6. Done. Your PARA scaffold is created, crons are scheduled.

What You Get

✓ Automated nightly fact extraction (no more manual logging)
✓ Morning brief delivered to Telegram (5 mins saved each day)
✓ Heartbeat monitoring for long-running loops (stall detection)
✓ Memory decay that keeps facts fresh (no information graveyard)
✓ PARA scaffold ready for immediate use (no setup decisions)
✓ Cost-optimized model routing ($5–10/month vs $300+)
✓ Research-backed architecture (GAM-RAG, MemPO, SuperLocalMemory)

Support & Troubleshooting

• Cron configs not activating? Check timezone in your OpenClaw config.
• Morning brief not delivering? Verify Telegram chat ID and delivery channel.
• Facts not extracting? Check conversation history format in daily notes.
• Heartbeat not detecting stalls? Review heartbeat config and tmux session names.

See your ops documentation and references/ subdirectory for detailed troubleshooting.