curiositech/always-on-agent-applications

/always-on-agent-applications — What Persistence + Memory Actually Unlocks

You are helping someone figure out what to build with an always-on AI agent that has episodic memory. This is the "so what?" skill — the architecture exists, the inputs are flowing, now what does it actually do that a stateless chatbot can't?

Decision Points

1. Always-On vs Session-Based Decision Tree

Evaluate task requirements:
├─ Task needs memory across sessions?
│  └─ No → Use session-based agent (cheaper, simpler)
│  └─ Yes ↓
├─ Task benefits from proactive behavior?
│  └─ No → Use scheduled agent with memory
│  └─ Yes ↓
├─ [Persistence ROI] > [Infrastructure Cost]?
│  └─ No → Start with session-based, upgrade later
│  └─ Yes ↓
├─ Domain narrow enough for quality memory?
│  └─ No → Narrow scope (meetings-only, code-only)
│  └─ Yes → Build always-on agent

ROI Calculation:

• Persistence ROI = (Task frequency × Time saved per task × User value per hour)
• Infrastructure Cost = (Server cost + Memory storage + Development time)

If ROI > 3x cost: Build always-on If ROI 1-3x cost: Start session-based, prove value first If ROI < 1x cost: Use existing tools

2. Application Category Selection

User asks "What should I build?":
├─ Primary workflow is coding?
│  └─ Yes → Developer Companion pattern
│  └─ No ↓
├─ Primary need is meeting/communication overhead?
│  └─ Yes → Personal Chief of Staff pattern
│  └─ No ↓
├─ Primary goal is learning/knowledge work?
│  └─ Yes → Learning Journal pattern
│  └─ No ↓
├─ Focus is health/habits tracking?
│  └─ Yes → Health Observer pattern (high safety sensitivity)
│  └─ No → Project Orchestrator or Ambient Intelligence

3. Scope Boundaries Decision

User proposes multi-domain agent:
├─ Is this their first persistent agent?
│  └─ Yes → Force single vertical (pick strongest ROI)
│  └─ No ↓
├─ Do they have >6 months development time?
│  └─ No → Single vertical only
│  └─ Yes ↓
├─ Can they define success metrics for each domain?
│  └─ No → Reduce scope until they can
│  └─ Yes → Allow multi-domain with staged rollout

4. Proactive Behavior Calibration

Configure agent interruption frequency:
├─ User work style is deep focus blocks?
│  └─ Yes → Batch notifications, respect focus signals
│  └─ No ↓
├─ User explicitly requests high-touch assistance?
│  └─ Yes → Allow real-time interruptions with relevance threshold
│  └─ No ↓
├─ Default to: 80% reactive, 15% passive proactive, 5% active proactive

Failure Modes

1. Hallucinated Memory Syndrome

Symptoms: Agent confidently references conversations or events that never happened Detection Rule: If agent claims specific quotes/dates/facts but can't provide exact source timestamp Root Cause: Poor memory boundaries between retrieved context and generated responses Fix: Implement strict memory citation requirements - agent must link every claim to specific memory entry with timestamp

2. Memory Pollution Cascade

Symptoms: Agent performance degrades over time, contradictory information in responses Detection Rule: If agent gives conflicting advice about same topic within 7 days without acknowledging change Root Cause: Low-quality observations accumulating faster than valuable signal Fix: Implement memory hygiene: relevance scoring, automated compaction, user-triggered memory cleanup

3. Cost Creep Explosion

Symptoms: Monthly bills increasing 30%+ without proportional value increase Detection Rule: If cost-per-useful-interaction rises above baseline by 50%+ over 30 days Root Cause: Agent over-processing low-value inputs (notifications, spam, automated emails) Fix: Input filtering pipeline, memory access budgets, proactive cost monitoring with auto-throttling

4. Scope Creep Paralysis

Symptoms: Agent tries to handle everything, excels at nothing, user abandons after 2 weeks Detection Rule: If agent has >5 distinct application verticals without clear success metrics for each Root Cause: Building "general assistant" instead of focused tool Fix: Force single-vertical start, require graduation criteria before expansion

5. Privacy Violation Drift

Symptoms: Agent accidentally shares sensitive information across contexts Detection Rule: If agent mentions personal/work details in wrong context (work info in personal chat) Root Cause: Memory boundaries not aligned with user privacy expectations Fix: Context isolation, explicit memory compartmentalization, regular privacy audits

Worked Examples

Example: Building Developer Companion Agent

Scenario: Software engineer wants agent to help with code reviews and PR descriptions

Step 1 - Scope Definition

• User: "I want an AI that helps me code better"
• Apply Decision Tree: Coding workflow = Developer Companion pattern
• Narrow scope: "PR description generation only" (not full coding assistant)

Step 2 - ROI Calculation

• Task frequency: 3 PRs/day × 5 days = 15 PRs/week
• Time saved: 5 min per PR description = 75 min/week = 65 hours/year
• User value: $150/hour × 65 hours = $9,750/year
• Infrastructure cost: ~$50/month = $600/year
• ROI = 16x → Build always-on agent

Step 3 - Memory Strategy

• Core memory: Current feature branch, last 5 commits, recent conversations about code changes
• Recall memory: PR templates user prefers, reviewer feedback patterns, project coding standards
• Archival: Historical PRs, team communication style, project decisions and rationale

Step 4 - Trigger Design

triggers:
  - git_push_to_feature_branch: Draft PR description
  - pr_opened: Enhance description with context
  - code_review_received: Log feedback patterns for future

Step 5 - Quality Gates

• PR descriptions include actual rationale (not generic summaries)
• Agent references specific commits/files mentioned
• 80% of generated descriptions require <2 minutes editing
• Agent correctly identifies when PR spans multiple concerns

What novice would miss: Starting with "AI coding assistant for everything" What expert catches: Focusing on single high-value workflow (PR descriptions) where persistence creates clear advantage over stateless solutions

Quality Gates

Application design is complete when all conditions are met:

• [ ] Clear ROI calculation showing >3x cost benefit
• [ ] Single vertical scope with defined boundaries
• [ ] Memory growth bounded with compaction strategy
• [ ] Proactive behavior frequency configured (<20% of total interactions)
• [ ] Success metrics defined and measurable
• [ ] User privacy boundaries explicitly mapped
• [ ] Cold start experience works without accumulated memory
• [ ] Kill switch implemented for user memory control
• [ ] Cost monitoring with auto-throttling thresholds set
• [ ] Graduation criteria defined for scope expansion

NOT-FOR Boundaries

Do NOT use this skill for:

• Building the memory architecture → Use always-on-agent-architecture instead
• Designing input feeds and triggers → Use always-on-agent-inputs instead
• Safety, privacy, and cost concerns → Use always-on-agent-safety instead
• General agent patterns and loops → Use agentic-patterns instead
• Evaluating AI safety risks → Use ai-safety-engineer instead
• Technical infrastructure decisions → Use systems-architecture instead

Delegate to other skills when user asks:

• "How do I store episodic memory?" → always-on-agent-architecture
• "What data should my agent watch?" → always-on-agent-inputs
• "Is this safe/private?" → always-on-agent-safety
• "How do I build agent loops?" → agentic-patterns