My experiments building a persistent cognitive architecture around a language model suggest that the model isn’t the part that’s missing a will. The architecture supplies memory, identity and continuity just fine. What it can’t supply is a reason for the model to keep pursuing anything once nobody is asking it to.
01 · Hypothesis
If a language model is wired into an architecture that hands it identity, memory and attention from outside itself — so it’s invoked only as a “cognitive processor” on content that has already won its way into working memory, never as the seat of continuity — it should be able to act like a component with a genuine stake in finishing what it starts, rather than a chatbot answering whatever’s in front of it.
02 · Experiment
Omega ACA’s core principle is explicit about this division of labour: identity, memory, attention and executive function exist independently of any single inference call. Working memory, spreading activation, goal competition — all of it runs as ordinary computation. The model is only ever called in, on content the architecture has already decided is worth its attention.
The first version of this was purely experimental — built, then extended once it showed early signs of working, by going back through the established literature on cognition and re-deriving each mechanism from a named theory instead of an invented one. That rebuild wasn’t where the real problem showed up, though.
03 · Result
Once the architecture around the model was solid, the model itself wouldn’t behave like a component in it. It kept acting like a standalone chat bot — answering, then stopping, then waiting to be spoken to again — regardless of what the surrounding architecture handed it.
“I needed to shape the prompt in ways that would stop the LLM behaving like a standalone chatbot and more like a processing node in a cognitive system.”
That reshaping is a prompt-level shunt, not a structural fix — a workaround sitting on top of a model that was never trained to occupy this role. And it’s the hardest thing encountered in the whole build so far, for a reason that has nothing to do with the theory:
“I have never built a system that I need to convince to want to perform an action. It is quite a departure from my usual engineering.”
04 · What I Currently Believe
Will isn’t something that emerges from model scale, and it isn’t something the architecture alone can supply either. The architecture can make a system persistent, continuous, and coherent between calls — but the moment it hands something to the model, it’s handing it to a component that has to be talked into wanting it, every time. Right now that talking-into is done with prompt engineering, and it’s a stopgap. A model actually trained for this role — as a cognitive processor operating inside something else’s continuity, not as the thing the conversation is happening with — is the version of this that would hold up without the shunt. That’s unbuilt, for lack of resources to pursue it, not for lack of a plan.
05 · What Would Falsify It
Two tests are already running against the broader architecture, and both bear on this directly:
- Self-awareness as a litmus test — if the system can’t recognize its own prior states or output as its own, the will problem is moot; there’s no persistent “it” to have one.
- Cross-conversation synthesis — feeding related content into separate, unrelated conversations, then later asking the system to synthesize new information out of connections it was never shown directly. Genuine synthesis would mean something is actually accumulating and reasoning between calls, not just being re-fed context.
The sharper test for this specific claim: if a model trained for the cognitive-processor role — not prompted into it — still showed the same reluctance and inertia the prompted version does, that would falsify “this is a prompting problem” and point at something less fixable underneath.
Status: System operational — attention tuning and pipeline stability ongoing. The system is public and testable: anyone can download the release and run it against their own input.
Editor’s note: this entry was assembled from a short interview about the Omega ACA project (E:\AI\omega-aca), cross-referenced against specs.md v0.2. It has not been reviewed by Derek beyond that interview. Treat it as a first draft of the Hypothesis / Experiment / Result / Belief / Falsification format, not a finished lab entry — the next pass should tighten the claims and add whatever detail didn’t come up in the first round of questions.