curiositech/always-on-agent-inputs

/always-on-agent-inputs — Feeding Context to a Persistent Agent

You are designing what an always-on agent sees, hears, and knows. The architecture skill handles where memory lives. This skill handles what goes into it — the raw signals, how they're structured, when the agent wakes up, and how to keep the context window honest.

---

DECISION POINTS

RETRIEVAL STRATEGY SELECTION

Agent Task Type → Retrieval Strategy

IF conversational/chat:
  ├─ Use recency-first (70% temporal, 30% semantic)
  ├─ Budget: 15K conversation history + 4K recent recall
  └─ Skip archival unless user asks "remember when..."

IF problem-solving/debugging:
  ├─ Use relevance-first (60% semantic, 40% temporal)
  ├─ Budget: 8K archival + 4K recall + 2K ambient
  └─ Include error patterns from past solutions

IF context-switch detected:
  ├─ Use frequency-based (what does user work on most?)
  ├─ Budget: 6K project context + 2K recent + 4K goals
  └─ Pull entity history for active files/people

IF scheduled trigger (morning briefing):
  ├─ Use structured agenda (calendar + tasks + updates)
  ├─ Budget: 4K calendar + 4K unfinished tasks + 2K changes
  └─ No conversation history needed

IF reactive trigger (CI failure, meeting starting):
  ├─ Use event-specific context loading
  ├─ Budget: 8K event context + 4K related history
  └─ Skip general conversation unless relevant

AMBIENT DATA FILTERING

Screen/Audio Observation → Filter Decision

IF same app + same text for >5 minutes:
  └─ DISCARD (no change, noise)

IF app switch detected:
  ├─ KEEP transition record
  └─ Score relevance of new app content

IF high-signal keywords found:
  ├─ Check against active projects/goals
  ├─ IF match score >0.6 → STORE to archival + recall
  ├─ IF match score 0.3-0.6 → STORE to recall only
  └─ IF match score <0.3 → DISCARD

IF technical error/exception visible:
  ├─ ALWAYS STORE (debugging context)
  └─ Tag with urgency level

IF calendar event mentioned:
  ├─ ALWAYS STORE (temporal anchor)
  └─ Cross-reference with actual calendar

TRIGGER URGENCY ASSIGNMENT

Event Type + Context → Urgency Level

IF CI failure:
  ├─ Production branch → ACTIVE (notify user)
  ├─ Feature branch + user currently coding → PASSIVE (mention when user next engages)
  └─ Old branch/PR → SILENT (log only)

IF meeting starting:
  ├─ <5 minutes → ACTIVE (interrupt with prep)
  ├─ 5-15 minutes → BADGE (show prep available)
  └─ >15 minutes → SILENT (prep in background)

IF code change detected:
  ├─ First commit on new branch → PASSIVE (note new work)
  ├─ Commit fixes previous error → PASSIVE (note resolution)
  └─ Regular commits → SILENT (track progress)

IF mentioned in Slack:
  ├─ Direct message → ACTIVE (respond needed)
  ├─ Channel mention + urgent keywords → ACTIVE
  └─ Channel mention, normal → BADGE (review when convenient)

---

FAILURE MODES

Schema Bloat

• Symptoms: Context window constantly maxed out, slow responses, high token costs despite filtering
• Root Cause: Observation schemas capture every field instead of just what the agent needs for reasoning
• Detection Rule: If average context size >50K tokens or observation storage >1MB/day, you have schema bloat
• Fix: Audit schemas every month. Remove fields that haven't influenced a decision in 30 days. Use summarization over raw storage.

Retrieval Thrashing

• Symptoms: Agent mentions outdated info, misses recent context, or retrieves irrelevant memories repeatedly
• Root Cause: Single-pass vector search without reranking, or temporal/semantic weights badly tuned
• Detection Rule: If >30% of retrieved context goes unused in responses, or user corrects agent about recent events
• Fix: Implement two-stage retrieval (broad search + rerank). A/B test different temporal/semantic weight ratios.

Trigger Spam

• Symptoms: User disables agent notifications, complains about interruptions, or agent gets ignored
• Root Cause: No urgency tiering, or thresholds set too low for active/interrupt triggers
• Detection Rule: If user dismisses >50% of active notifications, or avg time between user engagement >4 hours
• Fix: Default everything to silent. Let user explicitly configure what deserves active notification. Track dismissal rates.

Context Staleness

• Symptoms: Agent acts on outdated information, misses that user changed focus/projects
• Root Cause: No freshness decay on context, or retrieval weights don't factor in recency enough
• Detection Rule: If agent references info >7 days old when recent context exists, or mentions completed projects
• Fix: Add exponential decay to relevance scores. Weight recent observations 2-3x higher than archival.

Ambient Noise Flooding

• Symptoms: High costs, poor response quality, agent mentions irrelevant screen context
• Root Cause: Raw ambient data bypasses filtering, or relevance gates set too permissive
• Detection Rule: If ambient observations make up >30% of stored memories, or token costs >$10/day
• Fix: Tighten relevance scoring. Batch ambient observations in 5-minute windows. Require 2+ signals to confirm relevance.

---

WORKED EXAMPLES

Scenario: Billing Bug Investigation with Token Budget Constraints

User reports: "The billing calculation is wrong for enterprise customers."

Step 1: Context Budget Allocation (40K total)

System + Identity: 4K tokens (fixed)
Core Memory: 2K tokens (user preferences, active projects)
DECISION: Problem-solving task → use relevance-first retrieval
Allocated: 8K archival + 4K recall + 2K ambient + 15K conversation + 5K output buffer

Step 2: Retrieval Strategy Execution

# Broad semantic search for billing-related memories
archival_candidates = vector_search(
    query="billing calculation enterprise customer error bug",
    top_k=20,
    filters={"project": ["billing", "payments"]}
)

# Rerank by true relevance (not just embedding similarity)
reranked = rerank_results(archival_candidates, query, top_k=5)
# Result: Gets "enterprise discount logic bug from 3 weeks ago" 
# that pure vector search ranked #12

# Temporal search for recent billing work
recent_work = temporal_search(
    time_range="last_7_days",
    entity_filters=["billing", "enterprise"],
    max_results=3
)
# Result: User was debugging pricing tiers yesterday

Step 3: Ambient Context Integration

# Check what user is currently doing
screen_summary = get_ambient_summary(last_30_minutes)
# Result: "User has billing_calculator.py open, looking at 
# calculate_enterprise_discount() function"

# Budget check: 8K archival + 4K recent + 2K ambient = 14K (under budget)

Step 4: Fast vs Complete Trade-off Agent has two options:

• Fast retrieval (current): 14K context, focuses on recent bug patterns
• Complete context: Add git history, related PR discussions (+8K tokens, exceeds budget)

Decision: Stay with fast retrieval for initial response. If user asks for deeper investigation, trigger a follow-up with expanded context budget.

Agent Response Quality:

• Novice approach: Would dump all billing-related context, exceed budget, get confused response
• Expert approach: Curated 14K of most relevant context, mentions the specific enterprise discount bug from 3 weeks ago that vector search alone would have missed, stays under budget

---

QUALITY GATES

• [ ] Context window allocation explicitly budgeted across tiers (system/core/retrieved/conversation)
• [ ] Total per-request token count averages <40K for routine interactions
• [ ] Ambient observations filtered through relevance pipeline before storage
• [ ] Retrieval uses two-stage process (broad search + reranking)
• [ ] All triggers categorized by urgency (silent/badge/passive/active/interrupt)
• [ ] Observation schemas include timestamp, confidence score, and source
• [ ] Filter pipeline tested: <30% of retrieved context goes unused in responses
• [ ] User dismisses <20% of active notifications (trigger spam check)
• [ ] Agent references recent info (last 24h) when available vs defaulting to archival
• [ ] Storage growth rate <100MB/week per user (efficiency check)

---

NOT-FOR BOUNDARIES

Use /always-on-agent-inputs for:

• Deciding what data to capture from user environment
• Structuring triggers and observation schemas
• Context window budgeting and retrieval strategy selection
• Filtering ambient data before it reaches the agent

Do NOT use for:

• Memory architecture design (vector stores, databases, memory tiers) → use /always-on-agent-architecture
• Application and use case ideation → use /always-on-agent-applications
• Privacy and safety of captured data → use /always-on-agent-safety
• General prompt engineering techniques → use /prompt-engineer
• Agent reasoning patterns → use /agentic-patterns