Skip to main content

[Mind] technical tool: spoiler notepad

Feature RequestsMemoryAgents

An ability for the model to directly read and write into a notepad that the user can enable/disable but can’t open unless they pass a modal by confirming that they accept the spoilers it contains.

Let the model write there freely, and use the Venice RAG to read from it. 

It can be used to store plots features like game ending, game solutions, etc. 

1 comment

Log in to comment and vote

Comments1

  • Coffee Asteroid

    •

    Sep 22

    This notepad could be a tool, so we can choose to enable it for agents individually, and only that agent can read in their notepad. Each subagent would have their own space to write. That would do wonders for games and roleplay in general, but not just that. 

    Not so dissimilar to the memory/information feature we had with characters, BUT
    1- the LLM has agency over it,
    2- it works like a yaml file: so that the model can decide how to group data using separators,
    but is read line by line like an array of long strings: so that reading doesn’t crowd the contextual window when the content payload is added to our next prompt.
    3- we can disable it, read it, erase everything - so we keep control on our local data. 

    Detail:

    An agentic LLM use commands to freely write / read / delete notes (localStorage but like a table of long char,

    When the LLM uses a command to write, it chooses search tags for to the entry (command parameters: content, tags),
    and your system generate an id for it, add the note with the tags as metadata in localStoage.

    When the LLM use a grep-like command, or a clear command,
    for all notes that it wrote with specific tags or content,
    your system and use both the id and the tags to identify the note,
    the n add it to our prompt context, or erase it, according to the agent’s command. 

    Potential:

    That would give the LLM much more agency over its memory,
    and since memory changes a lot, the quality of any model’s answers would cap up to it’s max potential.
    It wouldn’t forget anything anymore unless we want it to,
    or unless our localStorage has reached its limit. 

    It would create strong narrative cohesion during much longer conversations, even for models that have a short context window, 

    Since it would also keep track of data, long scenarios would become playable in games and simulations, like TTRPG. 

    Better than Claude and Grok’s artifact system since it’s more precise and the user has more control over what the LLM write, entry by entry.