rinineaitools
patreon
How and why the memory manager works
🕑 Added 2023-05-23 07:01:58 +0000 UTCTo avoid any doubts, I will explain how the system works and why it achieves long-term memory.
The system provides you with a series of customizable text blocks for you to input information and memories.
That information is injected into the message input box every 4 messages, creating an automatic reminder for the bot to remember what is happening and the importance of your story or character. Therefore, it is ideal that during your role-playing session, you confirm what is happening and add new information to the memory manager to keep a record of the significant events that have occurred in the story.
You can save and load the set of memories in a file to continue the session at another time or to have a starting base (such as a detailed paragraph about the costume and description of your character).
Tips:
- If you load a memory file BEFORE writing to the bot for the first time (or after refreshing the page), the memories will be immediately injected when you start typing, reminding the bot of the current state.
- It is preferable not to introduce memories before they occur. This is important because if you do, the bot will interpret it as information you want to happen or general information, rather than as memories.
- In the extension options, there is a setting to test an alternative injection mode (via OOC, out of context), which can be more consistent and prevent the bot from reacting directly to that information. However, in this mode, the bot may be more prone to making small comments out of context.
- If you receive an out-of-place comment, swipe right! The AI model from character.ai is what it is and not perfect. Sometimes it says silly things, but generating a new response from the bot will provide an appropriate and coherent one based on the memories.
Comments
Fifth Vertex
Unfortunately this is an internal bug of the character.ai model that only they can fix.
goten will
chris, i have a question, does this prevent when we reach a certain point where the bots start to repreat themselves? here's an example: "with his evil smile...his evil smile..he did it..he smiled with his evil smile....* They go "...." with a word and start repeating it every time its their turn to reply, does memory manager prevent that?
Fifth Vertex
Reacting to keywords I don't think is very appropriate. In your own example you can be talking about "your outfit" or "the bot's outfit" and produce unwanted injections. There the possibilities for unwanted injections are almost endless. Considering that you can inject all the memories without problems of limitations, it is better to play it safe with everything. Regarding the dictionary format is something I am already working on. I am preparing a more detailed configuration menu to manage how the injection is done. Instead of choosing between narrative (*text*) or out of context (ooc Events occurred: "text", "text"), I am preparing a window to be able to have full control. This week I've seen the bots on their own doing things like [Thinking: I don't like this] [Not Thinking: I want to do that] So I'm going to prepare in that new window the possibility for the user to choose with which characters and how to enclose the memories. For example: (editable options in the new window) Opening of header: "[Memories:" Closing header: "]" Encapsulation per memory: N/A (empty) or "","" (enclosed by quotation marks separated by commas). etc. With 4 configurations to choose from a drop-down (to keep the current format of arterisks, occ, another new one for [Memories: "text","text","text"] and custom format (being customizable). Regarding your quote, I was talking about the default mode, which is by narration (arterisks). If you send a text between arterisks to the bot, it doesn't matter if you wrote it manually or if it was injected by the memory manager, the bot will react to it if it is a new event. On the other hand, if it is a event it already knows, it will ignore it but will remember it for when it is needed. That is why I say that it is advisable to let the events happen first before writing down the memories. But that does not change the fact that if you send a large paragraph with questions, the bot will assimilate the narrative but will not comment on it if the accompanying questions are not directly related to the narrative.
