Cognition across time
Attention decides what matters now. Temporal cognition decides how now relates to later.
A system that cannot represent time treats all decisions as immediate decisions, even when consequences unfold over minutes, weeks, or many sessions. It cannot distinguish between an important task with a month-long deadline and an important task with a five-minute deadline. It cannot recognize that the current moment is unusually dense with competing demands. It cannot organize past events into a sequence that reveals how a situation developed.
Temporal cognition introduces explicit representations of time as a cognitive dimension.
Planning horizons
The temporal engine represents five planning horizons that align with the goal hierarchy described in the goals article:
Micro: the timescale of the next action, measured in seconds. At this horizon, the system plans one step at a time, optimizing for the immediate next move given the current state.
Short: the timescale of a session, measured in minutes to hours. At this horizon, the system plans to complete a specific task within the current interaction.
Mid: the timescale of a project, measured in days to weeks. At this horizon, the system tracks progress toward a larger objective that spans many sessions.
Long: the timescale of a strategic arc, measured in months. At this horizon, the system maintains durable direction that should survive many project completions.
Meta: the timescale of the system's own development, measured in whatever units make sense for the deployment. At this horizon, the system reasons about how its own reasoning strategy is evolving.
The active planning horizon determines how far into the future each decision should consider, which memories are temporally relevant, and how urgency decays or grows as deadlines approach.
Urgency gradients and the micro-takeover
Urgency is not a binary flag. The temporal engine computes urgency gradients so that deadlines become more demanding as they approach, and opportunities decay as they recede.
Experiment 25 demonstrated a dramatic version of this: the temporal context flip. In the baseline condition (mid- and long-horizon tasks active, no immediate deadline), the active planning scale was "mid" and computational density was 0.300.
Adding a single micro-scale task (importance = 0.95, deadline five minutes away) instantly changed the active scale to "micro" and drove density to 1.000. One urgent item with a near deadline overrode the entire planning horizon.
This is the correct behavior. When an incident is imminent, incident response must take priority over the system's longer-horizon work. The temporal engine encodes this priority structure mechanically: micro-scale tasks with high importance and imminent deadlines will always drive the active scale to micro and saturate density.
The consequence for compute allocation is also important: at density 1.000, inferenceSteps rose from 10 to 16 and output compression tightened from 0.865 to 0.550. The system allocated more reasoning depth and tighter outputs to the high-density moment. This is the temporal engine coupling to the cognitive economics system described in the next article.
Subjective computational time
Inference depth need not be constant across all moments. A system that uses the same reasoning depth for a routine query and a high-stakes incident is misallocating its compute.
The temporal engine produces density signals that the cognitive economics system uses to adjust inference depth. High density (many competing demands, tight deadlines, high-stakes decisions) triggers deeper reasoning. Low density (routine tasks, no urgent deadlines, well-understood situations) allows shallow reasoning.
This is computational analogue to variable subjective time: people spend more mental effort on complex, high-stakes situations and less on routine ones. The architecture models this as a deliberate allocation policy rather than leaving it implicit in the reasoning process.
Episodic sequencing
Time is not just about urgency and planning horizons. It is also about order. Causes precede effects. Plans precede outcomes. Diagnoses precede treatments. A cognitive system that represents events only as a set of facts, without temporal ordering, cannot reason about causality, narrative, or how a situation unfolded.
The temporal engine maintains an episodic sequence: an ordered record of events within and across sessions, annotated with relative and absolute timestamps. This sequence is the raw material for causal inference (which event preceded which?) and narrative synthesis (how did this situation develop?). Both the causal engine and the narrative engine described later in this series depend on temporal ordering to do their work.
Why temporal cognition is essential for goals
Goals require temporal structure to be meaningful. A long-horizon goal is not just an important goal; it is a goal whose relevant timeframe spans many sessions. A micro goal is not just a specific goal; it is a goal that must be accomplished in the next few steps.
Without temporal representation, goals float free of the timescales they operate on. The system cannot distinguish "this must happen in the next five minutes" from "this should happen sometime in the next month." Both would be treated as goals with priority determined only by importance, and the five-minute deadline would not produce urgency unless urgency was encoded by hand.
Temporal cognition makes urgency automatic: a goal with an imminent deadline produces a rising urgency gradient that competes for attention as the deadline approaches. No manual urgency flag is needed; the temporal engine computes it from the deadline and the current time.