Every AI system in production waits. A request arrives, an answer goes back, and the system stops existing until the next request. That is a function, and it is a very useful one. It is also why none of them have a point of view.
There is a research project running right now that starts from the opposite assumption: that a mind is not a function returning a value but a process that does not stop, one holding interests, memory and a perspective that persists across time. It is worth looking at closely, because the architecture it arrives at is different from an agent framework in ways that matter.
Continuous Operation Is The Easy Half
The system never pauses. It takes a new turn every second or two with no prompt to start it and no answer to end it. That sounds like the hard part and it is not. A loop that never exits takes minutes to write.
Everything difficult happens inside a single turn. What decides the subject, where the knowledge comes from, how new information arrives mid thought, what happens when it hits something it does not know. Those are the questions that separate a running process from a thinking one, and they are all turn level problems.
Knowledge That Lives Outside The Weights
The model at the center is small and deliberately holds almost no knowledge. What it knows lives in an external memory it reads from and writes to as it thinks, structured so that every topic keeps its own summary and its own subtopics, each with its own record of everything encountered.
The reasoning is not about cost. Trained weights are frozen between runs and cannot change from one moment to the next. A memory can. Anything meant to live through time has to be changed by what happens to it, so the part that changes has to sit outside the part that is frozen.
The memory also scores what it holds and phases out the redundant material, the way people drop the details they no longer need. That is the piece teams building retrieval systems tend to skip. Retention without decay degrades retrieval quality, because everything competes for the same slots.
Choosing What To Think About Next
With nobody directing it, something has to set the agenda. Two drives do that here, one for disposition and one for motivation, feeding a scoring method that picks the next subject.
The more interesting half of that job is knowing when to stop. A model held on one line of thought for too long narrows, starts repeating, and eventually breaks down. Healthy sustained thinking depends on releasing a subject before that point, the way attention moves on by itself. Continuous operation is not survivable without it, which is why most of the build time went there rather than into the loop.
Writing Into A Thought Already In Progress
Retrieval is the simple half of the memory problem. The hard half is delivering what you retrieved into reasoning that is already underway without breaking it.
The approach taken is to work directly in the hidden state, the internal representation holding the reasoning before any words are formed, and write newly acquired knowledge into that live state. The information becomes part of what the system is thinking now rather than a fact handed over after the answer was already formed. This is described as one of the least settled parts of the design, which sounds right. It is also the part with the least prior art.
When it reaches a question it cannot answer from memory, it does not stall and does not ask a person. It sends worker agents to research the question, studies what they return, writes it into memory, and resumes where its curiosity left off.
What To Take From It
Whether consciousness can emerge from this arrangement is open, and the project says so plainly. It runs, shakily, on a MacBook Pro rather than rented GPUs, and the status report is honest about what is not working.
The useful part for anyone building agents is the separation. Frozen weights for capability, external memory for everything that changes, drives for what gets attention, and hidden state work for delivery. Most agent stacks collapse three of those into the context window and then spend their time managing the consequences.
The full write up, project stages and build log are at A Science Experiment in AI Consciousness.
Top comments (0)