Abstract Current neural architectures share weights across all users or sessions; the same model produces the same output regardless of who has been talking to it and for how long. We ask whether a single fixed architecture can, through lived experience alone, develop a genuinely different internal structure — not a different label, but a different competence — for different agents. We propose Self-Forming Agents, a continual-learning architecture built on three components: (i) a dual-timescale historical substrate in which a short-term decaying matrix R and a long-term pure-multiplicative matrix C accumulate experience in parallel; (ii) an encounter readout in which the current input draws out a response from the substrate rather than querying it; and (iii) a resonance gate that routes prediction through R and recall through C, firing C only when the input is familiar. A single multiplicative rule for encoding self-organizes a threshold and a saturation ceiling with no externally set parameters. We demonstrate, in a minimal model (N = 12), that: (a) the same architecture living different context streams individuates — each agent predicts its own lived contexts better than the other's (own-context advantage +0. 16, positive in 20/20 seeds, robust to harder tasks and to a 5x larger state space) ; (b) the long substrate C retains an early signature 1. 5× longer than the decaying short substrate R (n = 20 seeds) ; (c) recall from a 40 % fragment exceeds the raw fragment (0. 743 vs 0. 543) after correcting encoding to pure multiplication, a result that replicates on a second, structurally different task family; and (d) two separate learning signals — prediction error to R, reward to W (a valence layer) — cleanly dissociate all three functional layers (removing W collapses only valence, removing C collapses only recall, and prediction is nearly unaffected by removing either, with the small residual fully attributable to a deliberate, diagnosed design trade-off rather than an unresolved dependency). The architecture makes no claim that the resulting agents are conscious. It makes the engineering claim that the same model, living different experiences, can grow into a different agent that is good at different things — without external instruction. Version 3 update note: This version adds a polished, standalone implementation of the corearchitecture (kernelₚackage/) with documentation, error handling, andworked examples, intended for direct reuse. It also adds exploratoryscripts investigating multi-substrate routing (an open problem noted inthe preprint's Discussion) and its resolution, plus additional diagnosticexperiments. These additions extend and refine the codebase but are notyet reflected in the accompanying preprint text, which remains asoriginally submitted.
Kimiyasu Igarashi (Fri,) studied this question.