Description

During CI-204 testing of A-MEM link weight evolution, a self-referential edge was found in graph_edges:

id=2 | Anthropic Claude API → includes → Anthropic Claude API
retrieval_count=7, confidence=0.8

The entity extractor generates an edge where source_entity_id == target_entity_id. This happens when the LLM produces a relation like "Anthropic Claude API includes [capabilities]" and resolves both sides to the same entity.

Impact

• Semantically invalid: a node relating to itself has no information value
• Spreading activation: self-loops cause the node to accumulate activation from itself on each hop; with decay_lambda=0.85 and max_hops=3, this is bounded but adds noise
• Graph traversal / BFS: typically safe (visited-set deduplication) but adds an extraneous edge

Reproduction

Config: testing.toml with [memory.graph] enabled=true. Run several sessions with graph extraction active. Query:

SELECT e.id, s.name, t.name, e.relation 
FROM graph_edges e 
JOIN graph_entities s ON e.source_entity_id=s.id 
JOIN graph_entities t ON e.target_entity_id=t.id 
WHERE e.source_entity_id=e.target_entity_id;

Expected

Self-loop edges (source == target) should be rejected at the insertion boundary in GraphExtractor or the DB schema should enforce CHECK (source_entity_id != target_entity_id).

Severity

LOW — does not break functionality, but pollutes the knowledge graph and adds minor noise to spreading activation.