An engineering finding on why stateless LLM agents repeat choices, and how designing for memory is the solution, not better prompting.
The Illusion of Habit
In early testing, an autonomous marketing agent developed by Arcanamatic exhibited a peculiar habit. Tasked with selecting a psychological trigger for a series of campaigns, it chose ‘Authority Bias’ six times out of its first seven runs. It picked ‘Cognitive Ease’ five times. The immediate assumption was a model failure: a lack of creativity, or a collapse into a default state. The reality was more subtle, and points to a critical design principle for any system built on stateless API calls.
The agent wasn't being lazy. It was being logical. Each of the seven runs was a discrete, independent event. The model had no memory of its previous choices. An instruction to “be varied” is meaningless to a system that cannot see what came before; there is no baseline from which to vary.
Designing for Memory
The issue was not a prompting problem, but a system design problem. Two architectural changes were implemented to address this.
1. Contextual History
The system now feeds a list of the most recent selections back to the model as part of the prompt for the next decision. The agent is now state-aware, not because the underlying LLM has memory, but because its operating context now includes a history. It can see its recent tendency toward ‘Authority Bias’ and adjust its strategy accordingly if the prompt requires novelty.
2. Structural Constraints
The agent's choices are now also constrained by the marketing context it is operating in. A campaign is defined by its position in the funnel—brand, consideration, or conversion. Each stage has a pre-approved set of appropriate psychological triggers. A top-of-funnel brand piece, for instance, is now restricted to triggers like Reciprocity or The Halo Effect, preventing the model from selecting a high-urgency trigger like Scarcity where it would be incoherent.
When Repetition is Correct
This work uncovered a crucial distinction: the difference between a model defaulting and a model making the correct choice repeatedly. Several of the initial ‘Authority Bias’ selections were for technical articles aiming to prove an engineering point. For that context, appealing to authority is not just a valid choice; it is arguably the optimal one.
Forcing variety for its own sake would have produced a less effective, less coherent campaign. The goal is not to eliminate all repetition, but to ensure that when it occurs, it is an intentional and correct strategic choice. The engineering challenge is to give an autonomous agent enough context to tell the difference.
This finding from Arcanamatic’s development of its NeuraScene platform is a specific instance of a general rule for building with memoryless systems. True autonomy doesn't emerge from a single, brilliant prompt, but from the thoughtful design of the architecture in which the model operates.
Method note. This piece was built in NeuraScene's Advertising Mode using an explicit psychological frame:
- Persuasion trigger: Reciprocity
- Target affect: Flow State
- Gestalt principle: Law of Similarity
The post provides a specific engineering insight, an act of Reciprocity that builds trust with a technical audience. The content is structured to induce a Flow State, absorbing the reader in a complex problem. The Law of Similarity is the core of the problem itself: visually and conceptually, the agent's repeated choices appear similar, leading to a false inference of a unified 'habit' rather than discrete, identical outcomes.
