Letta agents, utilizing the GPT-4o-mini model, achieved a 74% accuracy score on the LoCoMo benchmark by simply storing conversation histories in files. This approach challenges the idea that specialized memory tools are necessary for effective memory management in AI agents. Instead, it highlights that the way agents manage context is more important than the specific retrieval methods employed. Without long-term memory, agents struggle with forgetting information and maintaining focus during complex tasks. MemGPT introduced a memory management system that organizes memory into layers, allowing agents to maintain significant context without being limited by fixed memory constraints. However, evaluating memory tools like Mem0 and LangMem has proven difficult because an agent's memory performance often relies on its ability to utilize these tools effectively. For instance, even if a memory tool theoretically outperforms others, poor agent training can hinder its effectiveness. Consequently, memory evaluations have largely focused on retrieval benchmarks rather than the broader context of agentic memory. Letta's recent experiments with its filesystem reveal that agents can use standard file operations to outperform specialized memory tools. This finding underscores agents' proficiency in managing their own queries and searching through data iteratively. The article also emphasizes that an agent's architecture and its underlying model play a crucial role in its memory capabilities. The Letta Memory Benchmark provides a consistent framework for comparing different models' memory management skills, moving beyond simple retrieval to assess dynamic memory interactions.