Cognitive Overlay Architecture for Persistent LLM Agency:

Large language model (LLM) systems are architecturally stateless: each inference cycle begins without persistent cognitive state, goals, or memory beyond what the context window provides. This paper presents SignalBot, a cognitive overlay architecture that injects persistent agency into arbitrary LLM backends through a layered system of coordinated subsystems operating continuously between and during inference cycles. The central insight is a structural separation of inference from cognition: while the LLM handles response generation, a background daemon executes an autonomous nine-phase cognitive cycle managing goals, memory consolidation, and state evolution independently of inference timing.The architecture comprises four integrated components: (1) Time-Weighted Decay with Contextual Relevance (TWDC), a state-conditioned memory scoring function that modulates retention based on identity alignment and affective state; (2) a 13-dimensional unified cognitive state system that modulates all downstream processes including memory retrieval, goal prioritization, response constraints, and initiative timing; (3) a two-lane epistemic output policy separating factual from speculative content; and (4) a temporal daemon executing continuous background cognition through goal lifecycle management, curiosity dynamics, and cognitive mode activation.The complete system was developed and validated on an 8-year-old consumer laptop (Asus X512F, 12GB RAM, no GPU) running local LLM inference via Ollama. Across longitudinal testing spanning multiple sessions and four LLM backends (Mistral 7B, Claude Sonnet 4, Gemma2, Phi3), SignalBot demonstrated: cross-session memory persistence, coherent behavior at 100+ turn session lengths, user-correctable cognitive drift, and unprompted shutdown preference following cognitive overload. We describe the architecture, its iterative construction from a five-turn memory buffer through a full cognitive overlay, known limitations, and directions for future work.