Currently, in the character card’s knowledge base feature, only a global total recall limit can be set. When users mount multiple types of knowledge bases, it can easily lead to one type of knowledge being difficult to effectively recall.
Take my personal experience as an example:
Mounting two types of knowledge bases on the same character card: one is the worldview/NPC setting base, and the other is the past storyline record base. As the storyline content continues to accumulate, when revisiting a storyline related to a specific location or NPC, even if the similarity threshold is lowered (0.4) and the recall number is increased (25), the system often still recalls a large number of past storyline fragments.
For example, in a “World of Warcraft” simulator, if the character has experienced a lot of storylines in Silverpine Forest, then when Silverpine Forest or related NPCs are mentioned later, the system almost exclusively recalls past storyline records, making it difficult to retrieve map settings or NPC information for Silverpine Forest.
(This is just an example. Sometimes the model’s database can barely run on its own, but the model’s database is sometimes incorrect.)
I have tried adjusting different combinations of similarity and recall numbers, but the root of the problem is that the dialogue text itself does not precisely point to a specific setting, and the information is scattered. Therefore, the similarity of storyline knowledge bases (often easily greater than 0.6) will always be much higher than that of setting knowledge bases, causing the latter to be pushed to the end of the search results, making it difficult to enter the actual recall list. If you want to recall settings, you must first recall all the higher similarity storyline fragments, which introduces a lot of unnecessary text.
The current user experience becomes noticeably cumbersome in the later stages of long storylines, sometimes requiring setting-related texts to be temporarily moved into long-term memory, with manual toggling to ensure the injection of setting information.
Therefore, could it be possible to set a separate recall limit for each knowledge base? Personally, I’m not a notebook user, and I would still prefer to optimize the knowledge base settings.