Persistent Individual Differences and Experience Transfer Depend on Distinct Mechanisms: Causal Intervention Evidence from a Simplified Agent Architecture

Abstract

If two agents begin from exactly the same state but experience different pasts, will they stillbehave differently after their current environments are made identical; and are the mechanismsthat maintain such differences also capable of reusing past experience in a new task? Weoperationalized these two questions separately and tested them with causal interventions in asimplified agent architecture that uses no language model and contains no learnable parameters.The persistence phase used a common-garden design: two agents sharing the same random seeddeveloped for 30 days in resource-rich and resource-poor environments, respectively, were thentransplanted into the same neutral environment, and behavioral differences were measured onlyduring the 30-day post-transplant window.Across 1,500 preregistered confirmation seeds that had never previously been run, theratio of across-history behavioral distance to the cross-seed, same-environment baseline was1.142, 95% CI [1.098, 1.183] (per-seed sign-permutation p = 1.0 × 10−4, dz = 0.20). Theeffect was small and strongly heterogeneous, with approximately 41% of per-seed differencespositive. Mechanistic interventions showed that deleting the entire episodic memory storeat transplant did not change the result, and deleting semantic memory did not meet thepreregistered primary maintenance criterion. Instead, persistent differences depended primarilyon a one-way ratchet at the trait level and a self-reinforcing action–trait positive-feedback loop.Targeted manipulation of feedback strength increased the transplant ratio from 1.021 to 1.575;across 500 jointly randomized parameter configurations, the feedback-gain parameter was moststrongly associated with persistence (Spearman ρ = +0.44).We then tested whether historical information could produce functional transfer in a frozenprobabilistic reversal-learning task. When history was read out through the trait pathway, theprimary endpoint in Experiment 027, the difference in restricted switch latency, was −0.0798trial (95% CI [−0.1632, −0.0035], p = 0.046): the effect was statistically detectable, but theentire interval lay within the prespecified ±1 trial practical-equivalence range. In Experiment028, expanding trait readout scope while holding the total coupling budget fixed produced nogain (G = −0.002, 95% CI [−0.031, +0.023]), and arm A did not stably replicate the weak 027signal in a new seed block. By contrast, Experiment 029 formed relational experience underindependent Stable/Volatile acquisition histories and fixed traits to isolate the memory pathway,yielding ∆C = −0.883, 95% CI [−1.016, −0.753]. When only the internal action–outcomerelations within experience entries were shuffled, the effect retention ratio fell to 1.06% (95% CIupper bound 0.110, below the preregistered threshold of 0.25); when same-condition experiencefrom a different random seed was used as the donor, the effect retained 104.6%. These results support a constrained separation of mechanistic roles: the proposition thathistory leaves persistent individual differences is not the same empirical propositionas the proposition that historical information can be reused in a new task. Theformer was maintained primarily by history-altered trait dynamics, whereas the latter, inthis task, depended on relational structure in retrievable experience. This is not a directwithin-experiment competition between traits and memory: Experiments 027–028 and 029 useddifferent historical sources, different confirmation seed blocks, and different primary endpoints.Although transfer in 029 was detectable, its confidence interval did not lie entirely beyondthe preregistered SESOI of 1 post-change error. The formal interpretation therefore remains:“detectable memory-mediated transfer, but functional significance not established.”