Milestone 3.4.5 — Memory Conflict Detection

Status: Planned
Goal: 3.4 — Memory extraction worker
Phase: 3 — Memory Engine and Operator Platform
Estimated effort: 3 days
ADR required: ADR-0037 — Conflict detection and resolution strategy

Why This Milestone Exists

AI agents accumulate memories over time. Memories can conflict: the user said "I prefer Python" in March and "I'm switching to Go" in June. The new memory contradicts the old one. Without conflict detection, both memories coexist and the agent gives contradictory advice.

Conflict detection is triggered during the write pipeline (Step 6). This milestone implements the detection algorithm and the automatic resolution strategies.

Conflict categories

Type	Detection	Resolution
Contradiction	High similarity (>0.85) + opposite semantic polarity	Mark old as `superseded`; write new as `active`
Supersession	High similarity (>0.92) + new information is a superset	Mark old as `superseded`; new is `active`
Duplicate	Identical content_hash	Skip write; return existing
Near-duplicate	Cosine > 0.92, different hash	Merge: write new combined memory, supersede both

Deliverables

Contradiction detection via secondary LLM call

Python

CONTRADICTION_PROMPT = """
Given two memory statements, determine if they contradict each other.
A contradiction means both cannot be true at the same time.
 
Memory A: {memory_a}
Memory B: {memory_b}
 
Respond with JSON: {"contradicts": true|false, "confidence": 0.0-1.0, "reason": "..."}
"""
 
async def check_contradiction(
    memory_a: str,
    memory_b: str,
    client: OpenAI,
) -> tuple[bool, float, str]:
    """
    Returns (contradicts, confidence, reason).
    Only called when cosine similarity > CONFLICT_THRESHOLD (0.85)
    to avoid LLM calls on unrelated memories.
    """
    ...

Resolution actions:

contradicts=True, confidence>0.8: Mark old memory superseded_by=new_memory_id. Create memory_relationship(type='contradicts'). Write new memory as active.
contradicts=True, confidence<0.8: Write new memory as pending_review. Create relationship type='contradicts'. Human operator resolves.
contradicts=False: Write normally.

Acceptance Criteria

Contradiction LLM call only made when cosine similarity > 0.85 (not on every write)
memory_relationships row created for every detected conflict
Old memory marked superseded when contradiction is high-confidence
Contradictory memories with low confidence written as pending_review (not auto-resolved)
ADR-0037 written covering detection thresholds and resolution strategies