Signal

Signal Memory ⚡

Status: 📋 Agent Guideline | Module: signal | Part of: Agent Brain

Conflict detection. The agent SHOULD call conflicts before storing new facts — this is a manual step, not automatic.

When to Run Signal

Signal is NOT automatic. The agent must explicitly call it:

1. Before storing a new fact: Run ./scripts/memory.sh conflicts "<content>" before add
2. On-demand: User asks "check for conflicts" or "anything inconsistent?"

# Before adding any new entry:
./scripts/memory.sh conflicts "User prefers Python for data work"

# If NO_CONFLICTS → proceed with add
# If POTENTIAL_CONFLICTS → ask user or supersede

How Conflict Detection Works

The engine filters out common stopwords (I, the, is, etc.) and compares meaningful words between the new content and existing entries. A potential conflict requires:

• At least 2 meaningful words overlapping
• The overlap covering at least 30% of the shorter text's meaningful words

This prevents false positives like "I like Python" vs "Python is a snake" (different context, only 1 meaningful word overlap after filtering "I", "is", "a").

Conflict Types

Direct Contradiction

Existing: "User prefers TypeScript"
New:      "User prefers Python"
→ Ask: "Previously you said you prefer TypeScript. Has that changed?"

Temporal Update

Existing: "Alex works at CompanyA"
New:      "Alex works at CompanyB"
→ Not a conflict — supersede the old entry
→ Run: ./scripts/memory.sh supersede <old_id> <new_id>

Context-Dependent

Existing: "Use short responses"
New:      "Give me detailed analysis"
→ Not a conflict — different contexts
→ Store both with context:
  ./scripts/memory.sh add preference "Short responses" user "style" "" "casual chat"
  ./scripts/memory.sh add preference "Detailed analysis" user "style" "" "research tasks"

Detection Flow

New content arrives
       │
       ▼
  conflicts <content>
       │
       ├── NO_CONFLICTS → proceed with store
       │
       └── POTENTIAL_CONFLICTS (with overlap %)
              │
              ├── Same topic, different claim? → Ask user
              ├── Same topic, newer info? → Supersede
              └── Different context? → Store both with context field

Response Templates

Contradiction Found

"I have something that might conflict with this:
 - Previously: [old claim]
 - Now: [new claim]
 Should I update, or are both true in different contexts?"

User Corrects You

"Got it, tracking that correction."
→ ./scripts/memory.sh correct <old_id> "<new_content>" "<reason>"

What Signal Does NOT Do

• Run automatically before stores (agent must call it manually)
• Monitor "tone shifts" (that's Vibe guidelines)
• Track confidence (that's Gauge guidelines)
• Run continuously in the background
• Detect "implicit" conflicts from silence or repeated questions

Integration

• Archive: Agent should call conflicts before add (not automatic)
• Gauge: Conflicts may warrant downgrading confidence to UNCERTAIN