How do you keep your AI agent from hallucinating over time?

I've been building my own AI agent for the past few months, and the one problem I keep coming back to is hallucinations.

At first I thought the answer was better prompts. Then I added memory. Then I redesigned the architecture, split long-term and short-term memory, added retrieval, planning, and verification. Every time I felt like I'd solved it, another weird failure showed up everytime.

Sometimes the agent remembers things perfectly and other times it confidently ignores its own memory or invents an answer when the information is already available. The frustrating part is that these failures aren't always predictable—they often happen after the system gets more capable, not less.

I'm curious how everyone else is approaching this issue.

If you're building AI agents, what's actually made the biggest difference? Better memory? RAG? Planning loops? Smaller context windows? Model choice? Multiple agents checking each other?

Or have you just accepted that hallucinations are something we'll have to live with for a while?

I'd love to hear what's worked in real projects, especially from people running agents in production.