Abstract
<p>Emotions are temporally structured: regret concerns what has happened, anxiety what will happen, joy what is happening now. Three strands of the active-inference literature each formalise one temporal face of affect: backward valence as the negative rate of change of variational free energy (Joffily and Coricelli 2013), present valence as reward prediction error (Pattisapu et al. 2025), and forward valence as the affective charge of policy revision (Hesp et al. 2021). Yet none alone spans the full temporal range, unifies these quantities into a taxonomy of emotion, or models how an agent actively selects the temporal orientation that dominates its own affect. We develop a single factored generative model in which a hidden temporal-frame state (past/present/future) redistributes precision across horizons, so that framing emerges from expected-free-energy minimisation. We motivate the architecture from the psychology of mental time travel, understood as the graded, cost-bounded regulation of how deeply memories and plans are unrolled, and show that a single depth regulator with two temporal faces reproduces the classical split between rumination (past-directed) and worry (future-directed). Affect here is a readout layer over a task-specific generative model, and the three channels are general operators on its inference dynamics. Driven by real experience-sampling sequences, the full model predicts next-moment affect out of sample better than linear autoregressive baselines, roughly doubling one-step skill where affect is volatile (held-out R² = 0.19 vs 0.09); the lead replicates in an independent second sample. On a static gamble it recovers the standard happiness equation (n = 14,803 held out) where each predecessor channel alone falls short, reaching that ceiling without expected value as a regressor: the forward channel supplies it by construction, a +30% gain over a reward-prediction-error channel. The same integration then supports capabilities the single-channel accounts lack: a latent temporal-frame state that tracks an independently measured symptom it is never shown (r = 0.17); counterfactual emotions with a behavioural signature a reward-only model cannot produce (regret drives choice-switching over and above one's own outcome, t = 10.4, replicated in a second dataset); and a mood layer whose diathesis-stress prediction holds in independent data, where a baseline vulnerability trait both lowers mood and amplifies affective stress-reactivity (vulnerability × stress, t = 4.2), a conjunction that in simulation collapses into a self-sustaining low-mood attractor. Throughout, we separate predictive claims (the channel integration) from generative ones (counterfactual and hedonic-asymmetry mechanisms).</p>