AI Agent Hallucination & Memory Leak Fix

Hallucination Fix

• Grounding: Require tool observation citation - "According to [tool X] result..."
• Ask agent to say "I don't know" - add to system prompt: If confidence <70%, say unsure.
• Reduce temperature to 0.1 for factual tasks
• Add verification step: second agent checks first agent's answer vs source

Memory Leak Fix

Symptoms: Agent gets slower, more confused after 10+ turns. Cause: never pruning. Fix:

Every 5 steps:
- Summarize last 5 turns into 1 sentence
- Keep only summary + last 2 observations
- Store full log to DB, not context

Tool We Use

LangFuse for memory tracing. Shows which memory chunk caused hallucination.