Definição
AI agent memory is the set of mechanisms that let an AI agent carry information beyond a single exchange: short-term memory held in the context window, episodic memory of past interactions, semantic memory retrieved from an external store, and procedural memory of how a task is performed. Unlike a model's parameters, which are fixed when training ends, agent memory is written and read at run time.
Porque importa
Agent memory is almost always built per agent and per vendor, so an organization running a dozen agents ends up with a dozen partial, unreconciled and ungoverned memories, none of which the organization actually controls. That is a fragmentation problem wearing the costume of a memory solution. Enterprise AI needs the opposite arrangement: one governed memory that every agent reads from and writes to, owned by the organization and portable across the models and platforms it rents.
De onde vem
An engineering pattern from the agent frameworks, borrowing its vocabulary from cognitive psychology. The Human Layer series positions it one tier down: agent memory is an implementation detail inside applications, while Organizational Memory is the layer beneath them all.
Ler os artigos de origem: The Human Layer Architecture · The Sovereign Memory Layer