Skip to content

Idea / Enhancement: use normalised vocabulary for facts #482

Description

@goughjo02

I should say that I have not actually used TractDecay yet. I have just been reading the repo and the documentation to understand how it works. If I misuderstand anything, please feel free to correct me.

While looking into how the holographic memory works, I noticed that the token vectors themselves do not contain semantic meaning. As I understand it, this means semantically similar facts with little shared vocabulary could receive a low similarity score. For example:

“The release failed because the credential was absent.”

“The production rollout broke due to a missing secret.”

This could weken recall. My first thought was that an embedding model could be used to compare semantically similar facts. However this is not really practical because TraceDecay is intended for basic claude / codex type users and embdedding models are not readily available to most people. So then I thought well it might not be so bad if the fact generation (the session relfector) normalised its vocabulary. There is a domain modelling skill here where i have seen this kind of thing done before. It enforces a shared vocabulary. Maybe something like that.

Again, I have not actualy used this yet. Just a thought I had while learning what you did here. Looks really cool. I definitely think that I will be trying it.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions