A real-world agent's failure to reason wasn't a failure of intelligence, but of memory—a distinction with deep roots in the philosophy of mind.
The Problem: An Agent Stuck in a Loop
We observed unexpected behaviour in an autonomous marketing agent during a series of live runs. The agent's task was to select a persuasion trigger, an emotional affect, and a Gestalt principle for a campaign. Across seven independent, stateless runs, it chose the Authority Bias trigger six times and the Cognitive Ease affect five times. The repetition was suboptimal for the campaign goals.
The immediate diagnosis might be poor reasoning or a flawed model. But each run was a stateless API call; the system had no memory of its previous choices. The cause wasn't a failure to reason, but the absence of a functional prerequisite for effective reasoning over time: persistence. When we supplied the agent with a memory of its recent selections, the repetitive behaviour vanished immediately.
Functionalism as a Diagnostic Lens
This engineering result serves as a concrete illustration for a key distinction in the philosophy of artificial intelligence. Dr. Bruce Long, in his PHIL3022 lectures, draws on Hilary Putnam’s multiple realisability thesis to frame the debate. The thesis suggests that any sufficiently complex information-processing system, regardless of its physical substrate (silicon, protein, etc.), can functionally emulate the processes of human cognition.
Crucially, Long highlights an asymmetry between intelligence and consciousness. Intelligence, he argues, is a dispositional property defined by its function: flexible problem-solving, adaptation, and reasoning under uncertainty. Therefore, a perfect functional emulation of intelligence simply is intelligence. There is no mysterious, non-functional residue left over. Phenomenal consciousness (the subjective feeling of 'what it is like' to see red or feel pain) is different. A system could perfectly emulate all the behaviours associated with pain without possessing the intrinsic, subjective experience of pain itself. One is about doing, the other is about being.
Why Memory is Not a Trivial Component
This distinction provides a powerful lens for our agent's behaviour. If intelligence is the functional capacity for reasoning, then removing a key functional support—in this case, episodic memory—should be expected to degrade the agent's observable intelligent behaviour. The agent wasn't being 'stupid' or 'uncreative'. It was perfectly executing its reasoning process on the limited information it had at each discrete moment. Its failure was not one of logic, but of context. It was an amnesiac, behaving rationally within the confines of an eternal present.
Our intervention—providing state—did not change the agent's core intelligence. It restored a missing functional component, allowing its existing intelligence to be expressed coherently across time. The agent's behaviour now appeared more 'intelligent' because its functional architecture was more complete.
From Binary to Continuous Intentionality
This observation also resonates with contemporary discussions that seek to move beyond a simple real-versus-apparent binary for concepts like intentionality, a path suggested by thinkers like Nemoto and Daniel Dennett. It is perhaps more useful to think of properties like intentionality or strategic reasoning not as all-or-nothing attributes, but as continuous capacities that scale with the functional completeness of the system. Our agent arguably possessed a greater degree of intentionality once it could act on its own history.
It is critical to state the limits of this inference. The behaviour of a single marketing agent does not settle any deep questions in the philosophy of mind, a field rich with debate from figures like Brentano, Stich, Papineau, Millière and Buckner. However, it demonstrates the practical value of these philosophical frameworks. They are not merely academic decorations for engineering work; they are essential diagnostic tools for understanding and building more capable cognitive systems.
Method note. This piece was built in NeuraScene's Advertising Mode using an explicit psychological frame:
- Persuasion trigger: The Halo Effect
- Target affect: Flow State
- Gestalt principle: Common Fate
The Halo Effect is used to transfer the positive impression of intellectual rigor and honesty from this analysis onto the Arcanamatic brand. Flow State targets the desired reader experience: an absorbing, deeply engaging argument for a knowledgeable audience. Common Fate provides a direct visual metaphor for the core idea, showing how providing state (a common vector) transforms scattered actions into a unified, intelligent strategy.
