Show HN: CoreTex – An Open-Source, Unix-like, biomimetic, flat-file AI Harness

I can say that many things are not yet fully wired up - HOWEVER - the true gain of function I see is in the following areas: - The Anterior Cingulate Cortex - I love enforcing consistency on memories. This is one tool in the toolbox of attacking semantic drift.

- The cerebellum: Even if it's not working well, the concept of turning previously executed code generated by an LLM into 0-cost reusable asset (which can be shared) is an important part of token economics

- The corpus collosum: the idea of right/left brain divide mapping to local LLMs vs cloud LLMs is an important cost saving and air gapping mechanism. We can have a local LLM watch dog off this with relative ease.

- The Blood Brain Barrier, Thymus, and Immune System - Doing AST scanning and using Audit hooks, combined with an efferent secret vault and a watchdog process on CoreTex are things we should have more of.

- Enteric System and Basal Ganglia - even though the BG isn't really working great yet, the idea that this system can repeat commands at 0-analysis-cost and then turn repeated actions into acron jobs based on changes to the file system (to avoid dependency on the host os) is another great quality of life feature.

- Vagus Nerve and Interoception - Creating a global token budget and aborting if things get too crazy is a good idea. I want the system to actively protect people from wasting money or denial of wallet attacks.

Those are a few! For more, you can see the biomimesis in docs. I could not have built this without following biology.

First, I want to support the same ontology and logic as Karpathy's LLM Wiki, but shifted towards data ownership and safety/consistency. I built in support for local LLMs for low cost / privacy, and use audit hooks + AST validation before writes for safety. Also the caching + indexing is local, and when coupled with anacron for repeated tasks and the Anterior Cingulate Cortex (blocking bad ops, inconsistencies, and infinite loop prevent), this will clean and monitor the wiki without needing to engage a cloud LLM or overspending if you do.

You don't have to switch - just give CoreTex your vault and see how it works on it. Keep in mind we are pre-alpha so it will be wonky.

What does CoreTex do currently to handle/prevent context degradation?

Marvin Minsky famously published a book in 1986 called "Society of Mind", in which he argued that people don't really have a "mind", but rather an enormous assortment of specialized subsystems that interoperate to produce the illusion of coherent consciousness. I think that this belief is what's needed to guide us forward towards AGI. IMHO, no further technological breakthroughs are necessary to achieve AGI; the remaining work is subsystem integration. I think that Daniel is hitting the nail on the head here with respect to the direction we should all be working in. After all, there already exists a machine capable of attaining human-level intelligence: the human brain. We know what the functions of many of its components are, and we know that those functions must have evolved for darn good reasons. Mimicking its architecture seems like a logical extension of Rosenblatt's original work of building artificial neural networks in the first place, sixty-odd years ago.