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.

Congrats on the launch, Daniel! Really cool approach. The flat-file architecture and sandboxed execution by default are smart design choices. The zero-token engram replay idea is clever too. Starred, excited to see where this goes.

Cannot wait to try it. What do you say to those who are leveraging 2nd brains like Karpathy’s LLM Wiki? What are the potential benefits to switching?

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

Hello HN! Decades long lurker, first time posting and here to hock my OSS project - https://github.com/mrdanielcasper/coretex.

I’ve spent the last ~30 days and 364 commits getting to the pre-alpha of CoreTex, an open source, Unix-inspired, biomimetic, text-only, portable, local, AI harness and knowledge engine. Breaking that down, my guiding principles and purpose were:

Purpose: I built CoreTex out of my personal need to organize my thoughts to be a better father, friend, and self-sufficient person. What began as the typical Obsidian-second-brain implementation quickly evolved into a quest to build a 1:1 analog of the brain itself. My hope is that this system will (eventually) help us organize our lives, achieve our goals, and build a community to change what’s happening to our world by working together. Maybe that’s simultaneously grandiose and naive, but I believe that machine intelligence can and should uplift all of us.

Principles:

- UNIX Philosophy: Flat files only like it’s the 1970s! State, IPC, data, queues, are files with concurrency guarded by OS-level locks. No databases here except an ephemeral SQLite FTS5 DB for performance. The payoff is threefold: The user and the AI can know all and see all (including its own source and state). CoreTex works with operators, redirect, pipe, etc. making it highly composable. Development velocity - very few things get in the way of adding features.

- Biomimesis: Why reinvent what nature already optimized? I followed evolutionary pathways to build the 39 biomimetic modules of CoreTex. Even the sensory organs (Sense) are decoupled as their own package and connected through the “Spine” which acts as a HAL. The payoff: 1) Procedural “Muscle Memory” (experimental) - the cerebellum creates engrams of repeated code, e.g. setting up a FastAPI + React app, so you can 0-cost repeat it later. It’s DRY for $. Engrams will be sharable over the exocortex via MCP/ACP. 2) “Self-healing” modules like the microglia intercept CLI errors at low cost. 3) A 5-tier memory system: Working Memory (token aware compression), Short Term Memory (FTS5 + BM25 ranking + snippet truncation), Relation Knowledge (maps dependencies to a serialized graph, protected by the ACC circuit breaker), Episodic Memory (thread-safe WAL JSONL ledger of actions), and Long Term Memory (rotates logs during sleep to extract key memories for alignment).

- Shift Left & Token Economics: Why don’t we shift left ALL THE THINGS? If we’re going to use this to help survive the Age of AI, we’re going to need 0-cost whenever possible. Local LLM support. Enforceable daily token budgets. Prompt caching, env guillotines, sliding context windows, transducers and tool truncation.

- Performance: No vectors in my lookups! FTS5 + BM25 + Ripgrep for text search, semantic compression of large files, Content Addressable Storage against a rapid hashing ledger for O(1) lookups, and deterministic transduction of CLI outputs at zero cost.

- Security: Safe by default, operating only in Cognitive mode - you have to turn on code gen. I resisted containers/microvms at first and attempted to make the most secure sandbox on the host OS possible. I realized my hubris after spending hours building a parent Watchdog daemon, switched to Deno + WASM, and eventually caved to support ephemeral Docker containers. Example - CoreTex’s Defense against Mini Shai Hulud: Code execution is set to false by default. When enabled, it runs in Pyodide and Deno where only secured CDNs are whitelisted over port 443. Dynamic path proxies prevent traversing unsafe paths or symlinks. (Where CoreTex loses: DNS exfiltration, because Deno hands DNS resolution back to the host OS where a worm could use subdomains).

The Current State: It will be rough using the pre-alpha build. Expect it to be “allergic” and “fearful” of your commands (thanks amygdala!). Expect to tweak agents.yaml. Expect to tweak your tasks. Expect a surprise token cost (<20k) for a simple task or API key issues. It may not even install properly. But use ./ctx daydream and tell me that it’s not something special to see a program suggest improving its own daemon? Or when I asked it to take my notes and make a GTM + website (autopoiesis) that it gave me an output that was a solid B? I feel like the bones of something are here.

Quo Vadis: I intend for CoreTex to be the freest, cheapest, and safest personal control plane + always on daemon, interoperable with every major AI system (Pi, Hermes, OpenClaw, OpenHuman, OpenWhatever) coupled with a robust ecosystem. That includes allowing CoreTex to connect to other instances (via the exoreceptor and exocortex modules) to share knowledge, code, memories, and compute in a gift economy. Ultimately, I want to give everyone a personal and free system that is totally aligned to your goals and your self, allowing you to observe and react to the world even as you sleep.

So with all that said, rip it to shreds! Every time I’ve gone through the crucible with my friends, CoreTex has come back stronger. Life…finds a way (even if it’s artificial).

P.S. I developed this on my Windows machine by arguing and co-creating every module, one at a time, with Gemini using my 100 Pro-Prompts-a-Day budget. Maximum portability. Minimal dependencies. No databases. And, most importantly, No items, Fox only, Final Destination.

Developing software w/ AI, It can be all to tempting to spend my time and resources on the AI system and tools itself, rather than the work we could be pointing it at.

What are you most excited / looking forward to using CoreTex for? Is there an immediate use case to try with it, that you think its uniquely capable for?

How much of the biomimetic structure is metaphor vs. architecture that materially changes behavior or performance?