February 2025 Summaries

AI agents begin each session without inherent continuity, so persistent memory must be deliberately designed, with human cognition offering a useful model. Human memory includes fleeting sensory impressions, short-term or working memory for actively holding and manipulating information, and long-term memory divided into episodic memories of specific events, semantic knowledge of general facts and preferences, and procedural knowledge of learned skills. In agents, raw input buffers loosely resemble sensory memory, while context windows and session-scoped information serve as working memory but disappear after a task or conversation. Long-term agent memory requires external storage and selective retrieval: episodic records can preserve outcomes from past interactions, semantic memory can capture recurring user preferences or domain knowledge, and procedural memory can retain reliable workflows and task sequences. The text argues that effective agent continuity depends not simply on saving all conversations, but on distinguishing these memory types and retrieving the most relevant information when needed, an approach offered by systems such as Mem0.