1/ I ran Claude Code for six months. Then I ran Codex. Then I ran both at the same time, hoping the combined cost would finally justify the output. It did not. The problem was not the models. Claude is excellent. The frontier models are the best available. The problem was memory. Every session starts from zero. The agent does not remember what you taught it last week, what worked, what failed, or why a decision was made. I wanted an agent that gets smarter over time because it remembers what it learns. 🧵

Full article and thread up. x.com/guideboardlabs/status/…

10/ The third theory is deterministic governance. I do not want a language model deciding whether governance rules should apply. Rules should be rules. Models should be models. Keep those responsibilities separate.

9/ The second theory is filesystem-first. Not Redis. Not Postgres. Not hidden internal state. Files. You can inspect them, version them, search them, back them up, and read them with grep. The simplest thing is often the most durable thing.

8/ The IDE layer is Cammander: browser-based, mobile-first, terminal, chat, git, vault access, project management. The interesting part is that Cammander was built using Hermes and OverCR. The stack built its own interface. That became a forcing function: if the tools are not good enough to build their own UI, they are not good enough for real work. (WIP) github.com/GuideboardLabs/ca…

7/ This separation ended up being one of the most important design decisions. Let the agent be creative. Let the substrate be rigid. One generates possibilities. The other enforces constraints. Different machines. Different jobs.

6/ Hermes is the runtime. It listens, plans, delegates, executes, and manages memory. Under Hermes sits OverCR. OverCR does not think. It enforces. Every action passes through a deterministic state machine. Every transition is logged. Every approval is auditable. Every execution is traceable. The agent is probabilistic. The governance is not.

5/ A lot of people assume this means RAG. It does not. No vector database. No embeddings. No similarity search. Facts are written by the agent, verified by the operator, and retrieved through deterministic search. Context is constructed. It is not guessed. Context Accumulation Generation.

4/ Current numbers: • 33 indexed notes • 436 facts • 32 domains The value of the system increases every time it is used. Most agents consume context. This one accumulates it.

3/ The vault is the important part. Most agent systems are a model talking to a tool. Prompt → Tool → Result. That works until the next session. My solution was a filesystem-backed knowledge vault. Every useful discovery gets written into structured markdown. Every new task begins by reading relevant facts before doing any work. The agent I run today has more context than the agent I ran yesterday. That compounds.

2/ So I built a stack around that idea: • @NousResearch Hermes Agent (runtime) • OverCR (governance substrate) • Cammander (browser IDE) • @obsdmd Vault (long-term memory) Plus three autonomous workers that maintain the system. Everything is open source. Everything runs on my own hardware. No per-seat licenses. No API key required to open the IDE. github.com/GuideboardLabs/ov…

You spent years fearmongering about safety to build the narrative. You lobbied for the regulations that would only hurt everyone else. And now you got exactly what you asked for.. the government doing your dirty work for you, locking your model behind a gate only you can open. This was never about safety. It was about the moat. I cut you out of my stack two weeks ago. No Claude Code, no API keys, no dependency on your roadmap or your regulatory strategy. Running Hermes OverCR Cammander on my own hardware. The agent remembers what it learned. The stack gets smarter every run. No one can flip a switch and take it away. Full writeup on the stack coming later.

Read this

Bro asked Fable 5 how many days have the letter D It hit him with the Wednesday has three It's over for us💀

What's crazy to me is that Fable is blocked from life sciences broadly, nerfed even if you get passed the classifiers and filter level blocks. The whole point of AGI/ASI is to cure all diseases. Everything else is just nice to haves. But Anthropic wants to close off that path. I think Anthropic might be the worst company on the planet.

3/ the plans, the decisions, the architecture reasoning…. vanishes. The next bottleneck in AI engineering is not model capability. It is continuity. The model is smart enough to build anything. It just has to remember what it was building tomorrow.

2/ Every capability upgrade widens the gap between what the model can do in one session and what the agent remembers across sessions. Fable 5 can run for days. Stripe migrated 50M lines in one day. That output is incredible. And when the session ends, all of that context….