Memory is not one switch
A chatbot may use recent messages as context, a separate saved-memory feature, a profile or character definition, or some combination of them. These mechanisms behave differently. Recent context can fall out of view as a chat grows; saved memory may be editable, limited, or disabled. The exact behavior depends on the provider and model, so treat any “remembers everything” claim as something to verify rather than a guarantee.
Build a low-risk recall test
Choose a fictional, non-sensitive fact—such as the color of a made-up café—and mention it once. After some conversation, ask a natural question that makes the detail relevant. Repeat in a new session only if the service presents memory as cross-session. Record whether the answer is correct, whether you had to restate the fact, and which memory setting was enabled. Do not test with your real address, health details, passwords, or another person’s private information.
- Find the provider’s current description of chat history, memory, retention, and deletion controls.
- Check whether you can view, correct, remove, or disable saved facts.
- Keep a short reference for fictional names, relationships, setting, and agreed tone.
- After a contradiction, restate the relevant canon in one sentence instead of adding a long recap.
Make consistency recoverable
Store your own concise character notes locally if continuity matters, and paste only the parts needed to resume a scene. Separate stable facts (“the character is a night-shift botanist”) from temporary scene state (“they are at the station tonight”). This makes corrections easier and reduces accidental prompt clutter. If memory controls are absent or unclear, assume persistence is uncertain and avoid putting sensitive material into chats. Provider-dependent details—including what deletion removes and whether settings apply to existing conversations—should be checked in current account documentation.
