Persistent memory is storage that survives application restarts and sessions, so an AI system can retain knowledge and context over time rather than losing everything when a process ends. It is what allows continuity: an assistant that still remembers your preferences tomorrow, or an agent that builds up knowledge across many runs.
This contrasts with the ephemeral nature of a language model’s context window, which holds information only for a single interaction and is gone the moment the call completes. Persistent memory lives outside the model, most often in a vector database that durably stores embeddings of past conversations, documents, and facts, ready to be retrieved whenever they become relevant again.
For agents and long-running assistants, persistent memory is essential to behaving coherently over time. By embedding new information as it arrives and retrieving relevant pieces on demand, the system gives a fundamentally stateless model the appearance of lasting memory. Effective persistent memory usually pairs durable storage with policies for what to keep, summarise, and prioritise, so the accumulated history stays useful rather than simply growing without bound.