Memory Coherence Breakdown

Issue: Retrieved Memories Contradict Each Other

Frequency: Common

Symptoms

• Agent states contradictory facts
• Different sessions have conflicting information
• User receives inconsistent responses
• Agent acknowledges contradictions but can’t resolve
• Memory contains multiple versions of truth

Root Cause Over time, memory accumulates conflicting information: corrections, updates, different contexts. Without coherence management, retrieval may return contradictory memories. The agent then produces confused responses or arbitrarily picks one version.

Example

Memories retrieved:
- "User's favorite color is blue" (2024-01)
- "User said they hate blue" (2024-06)
- "User prefers blue for work, red for personal" (2024-08)

Query: "What color should I use?"

Without coherence management:
Agent: "You like blue... but also hate blue... 
        but prefer blue for work..."
User: Confused and frustrated

With coherence management:
Agent: "For work contexts, you've said you prefer blue.
        For personal use, you prefer red. Which context
        is this for?"

Contributing Factors

• No conflict detection
• All versions stored without reconciliation
• No source/context tracking
• Updates append instead of replace
• No truth arbitration logic
• Missing provenance metadata

Eval Recipes

Test Cases

Metrics

Mitigation Strategies

Prevention

1. Conflict detection: Identify contradictory memories
2. Source tracking: Record where each fact came from
3. Version management: Updates replace, not append
4. Context tagging: Tag facts with applicable context
5. Resolution logic: Rules for which version wins
6. User confirmation: Ask user to resolve conflicts

Coherence Framework

On retrieval:
1. Fetch candidate memories
2. Detect conflicts (semantic similarity + contradiction)
3. Resolve conflicts:
   - Same context: Use most recent
   - Different contexts: Keep both, tag context
   - Uncertain: Ask user
4. Return coherent set

Production Signals

Key Metrics

Alerts

References

• Knowledge Base Consistency
• Belief Revision in AI