Unveiling Privacy Risks in LLM Agent Memory