i'm working on a programming language that eliminates shared memory complexity and performance penalties of using heaps and garbage collection. the issue of badly designed data access for tools to improve agents maintaining relevant context over broader and longer lasting context - giving it more memory, is a problem i intend to address by building my own agent.

it also dovetails into another project i'm working on, which is a semantic lattice algorithm that can replace a lot of the memory and context functions that LLMs have natively. i have a plan to build a small (maybe up to 3b parameters) that is taught primarily to reason and understand language, to a sufficient degree that it can call on the lattice for populating the context with relevant information. i've seen a few projects that attempt to create similar kinds of tooling but the integration was not good in the ones i have tried so far. and the other thing is i hate TUIs because they make copy-paste difficult. i usually run a general chat session in one window, alongside my coding agent sessions, and use the cleaner, smaller chat context to help with revision of planning, as well as journalling. so, the chat agent i have in mind will function as a smart journal as well as having a semantically structured access to broader, often time-related context that i would otherwise have to repeatedly explain key points for the agent to respond properly. things like - what i'm doing, where i am, who i am talking to, the state of my health, etc etc.