Long-running agents are broken. They forget critical facts mid-task, lose reasoning coherence across sessions, and degrade as interactions accumulate. The problem: we treat agents like stateless APIs, not stateful systems. Context windows are expanding (1M tokens), yet agents still struggle with context collapse, information loss, and irreversible decisions. A customer support agent processes 50 interactions but can't remember the customer's stated preference from conversation #3. A research agent reads 200 papers but loses track of critical early citations when new evidence emerges. The frontier is designing memory systems that span fixed-size context + external persistence + intelligent retrieval. Recent breakthroughs (ReMemR1, Anthropic's context engineering, MemGPT patterns) show how agents can maintain coherence over hundreds of thousands of tokens. This talk explores: The memory problem (context rot, hallucination, forward-only processing, irreversibility). Emerging architectures (ReMemR1: +6-10% accuracy, MemGPT core blocks, context engineering, sleep-time compute). Production challenges (consistency models, eviction, retrieval semantics). Case studies (multi-session agents that improve). This talk bridges academic research to production practice, giving engineers the frameworks to build agents that actually learn and remember.

Long-running agents forget critical facts, lose reasoning coherence, and degrade as interactions accumulate. Context windows expand (1M tokens) yet agents struggle with context collapse. Recent breakthroughs (ReMemR1, Anthropic context engineering, MemGPT patterns) show how agents maintain coherence over hundreds of thousands of tokens. This talk explores the memory problem, four emerging architectures, production challenges, and case studies of agents that improve with every interaction.