Amnesia Is All You Need

A developer building a GraphRAG system (Senegraph) for Stoic philosophy discovered that LLM agents with extensive domain knowledge in their training data would skip tool calls and hallucinate citations instead. Despite strong instructional prompts demanding tool use, the Seneca agent fabricated quotes 50%+ of the time. The breakthrough came from reframing the agent's identity: instead of telling it 'you must search,' the system prompt told it 'your memory is unreliable — you have amnesia.' This situational framing dropped fabrication to zero on first turns, because it reframes an actual LLM limitation as an intrinsic identity trait the agent cannot reason away. The post also covers a frontend parser that catches any fabricated citations and triggers a retry, ensuring users only ever see graph-backed quotes.