Choose what belongs in memory
Our conversation tools include pinned memories for details you want to retain as context. Choose a concise fact rather than a long transcript. Keep the participants clear, particularly when your persona and the AI character have different roles.
My character is Alex, an adult cartographer. Alex has not opened the sealed letter. The mountain bridge is closed until morning.
These are reusable scene facts. A request to answer the next question in two sentences belongs in a reply instruction instead.
Use instructions for the way replies are written
Custom instructions can set a preference such as brief dialogue, one question per reply or leaving your character's decisions to you. Describe the behaviour you want without repeating every event in the story.
Compare the wording with the character definition. An instruction to give very short replies conflicts with a request for several detailed paragraphs. Revise one preference and try the next exchange before changing everything.
Keep a changing story consistent
Review retained facts when the scene changes. If the bridge reopens or the letter is read, remove or revise the old detail so it no longer competes with the current story. A brief recap can help when you return after a break.
- Identify the fact the reply missed.
- Check the character definition and retained details for a contradiction.
- Restate the current fact in one sentence.
- Continue with one immediate question or action.
Keeping a detail as context is not a promise of perfect recall in every response. Save any writing you want to keep before resetting or deleting a conversation.
Keep personal information out of a story note
Use fictional scene details, not a password, identity document or payment information. A memory control is part of a conversation tool, not a secure document vault.
For writing examples, use chat prompts. For audience and data handling, review conversation privacy. If a message fails to send rather than losing a scene detail, use technical checks.