LLMs have the memory of a goldfish.

Every conversation starts from zero. They don't remember you, your preferences, or what you discussed yesterday.

This is one of the most important unsolved problems in AI, and solving it (even partially) is enormously valuable.

The approaches that work in practice: RAG for external knowledge (the model retrieves relevant context). Conversation buffers for short-term memory (keeping recent messages). Summary memory that compresses long histories into concise recaps. Entity memory that tracks specific people, projects, and things mentioned. Long-term vector storage that embeds and retrieves past conversations.

More sophisticated approaches are emerging. MemGPT lets the model manage its own memory like an operating system. Zep provides long-term memory as a service. LangGraph checkpointing saves agent state.

The chatbot that remembers your name, your preferences, your past conversations — that's not magic. That's careful memory engineering. And the difference in user experience is dramatic.

I've built chat systems with and without memory. The ones with memory have 3-4x higher user retention. People don't want to re-explain themselves every time.

If you want an impactful project, build a chat application with genuine long-term memory. It's harder than it looks and more rewarding than you'd expect.