Reviewing and correcting memory

Enable memory only where you need it#

Memory stores preferences and events for later conversations. It is separate from conversation history and the current scene. It is off by default. In the chat information panel, open “Remembered information” (覚えていること) and enable “Remember this conversation” (会話を覚える) to use automatic memory proposals, retrieval, and prompt injection.

1. Check the selected character and saved persona revision.
2. Enable the conversation's memory switch.
3. Review stored items; use “Add something to remember” (覚えてほしいことを追加) for a manual entry if needed.
4. Turn it off for conversations that do not need it. While off, automatic memory saving and retrieved-memory injection are disabled.

Turning memory off does not delete saved memories or conversation history. A model may still refer to information present in the visible history. Treat memory use and history management as separate concerns. See Data and privacy for data handling.

Check which conversations share memory#

The normal sharing scope combines the character ID and saved persona revision. Other conversations using that combination can retrieve the same memories. Free chat without a character keeps memory within that session. Experiments use their own independent memory scope.

For example, preferences saved with “Rio + traveler persona, revision 1” are not automatically carried into “Rio + the same persona, revision 2.” If memories appear missing after editing a persona, check the selected revision first. Migration across persona revisions is not automatic.

See Personas and worlds for the distinction between persona, setting, and scene. Editing a character or persona also does not automatically switch an ongoing conversation to its latest revision.

Write concise entries and correct errors directly#

Memory categories are preferences and facts, events and promises, and relationships. Keep each entry focused, with a clear subject and a short factual statement.

Use “Correct” (訂正) to edit an inaccurate entry, or “Forget” (忘れる) to remove it. Model proposals with weak evidence remain marked as pending review and unused in conversation. They are not injected until corrected. Decide whether to replace them with accurate information or remove them.

A chat reply such as “That's wrong” does not establish that the stored item was corrected. Check its text in the list afterward. This also does not erase earlier conversation text or exported backups, so inspect those before sharing.

Inspect storage, retrieval, and apparent use separately#

“Stored,” “retrieval candidate,” “injected into the prompt,” and “apparently used in the reply” are different stages. Memory Inspector in Researcher mode shows provenance, time, revision, candidates, injection, and estimates. If a saved fact is not recalled, inspect retrieval and injection before assessing the model's reply.

The public implementation retrieves memory using character bigrams. It does not guarantee semantic understanding or automatic resolution of contradictions. Apparent use in a reply is also inferred from signals such as text matches. Check important promises directly in the memory list and look for invented details.

Predefined character knowledge belongs in Lore under character settings. Return to the documentation index for all guides.