Persistent memory is information an AI system keeps after a session ends and can use in later sessions without being re-supplied.
Ordinary chat memory is the context window: it lasts as long as the conversation and is gone when the tab closes. Persistent memory is written somewhere durable, such as a file, a database or a hosted memory service, and read back later. The write can be automatic (the model extracts facts as you talk) or explicit (you or the agent save a checkpoint). The read usually happens at session start or just before answering a question.
Persistence raises questions that session memory does not. Who owns the data? Can it be exported? Can it be edited or deleted? Does it survive a vendor shutting down? ChatGPT memory, for example, is free but capped, auto-rewritten since June 2026, and not included in data export (see notis.ai). Files on your own disk answer those questions simply: you own them, you can copy them, and no one else can switch them off.
How Central Brain handles persistent memory
memory_recall in the next, from any connected tool.