A Time of Their Own: Agent-Held Temporal Evidence for Persistent AI Agents

Persistent artificial agents are ordinarily anchored in time from outside, and whatever timing they learn internally remains part of a state that can be saved and restored. When that state is restored, every internal clock and record is restored with it: from its restored state alone, an agent cannot tell whether it has continued without interruption or been rolled back. Persistence, continuity, and identity of such agents have recently been recognised as open problems, and mechanisms that supply evidence of another kind exist elsewhere — in hardware security, cryptography, and the neuroscience of timing — but they serve the platform, or the guard checks of protected modules, not an agent’s predictive model of itself. We classify the temporal evidence available to an agent — of order, duration, sequential computation, continuity, and its own change — by who can write, reset, forge, or copy it and by whether its self-model can read it, and we distinguish continuity claims adjudicated by the platform from raw evidence interpreted by the agent, treating their equivalence as a legitimate negative result. Our hypothesis is conditional: knowing how old one’s knowledge is, detecting suspension, restoration, or copying, and ordering, revoking, and anticipating one’s own changes benefit from authenticated evidence that is not restored with the agent’s state, chiefly when external records are absent, delayed, or untrusted; with complete, trusted logs, replay reconstructs any deterministic agent, and an internal alternative can only be cheaper. Physical irreversibility is one route to such evidence, not the argument for it. We separate two roles: a non-restorable time reference with a state-bound operation counter detects and distinguishes suspension and restoration, whereas an index of accumulated experience-driven change, read against that reference, serves the agent’s model of its own adaptation. A completed exploratory study supplies the evaluation method and shows why local update statistics must be excluded before traces are read as accumulated history. Each proposed test is paired with a conventional competitor that can close it, and access to evidence of restoration confers no control over it. We make no claim about consciousness or moral status.Version 1.1 cites prior and related work omitted from version 1.0 (Raell 2026a, b; Louck 2026; Su et al. 2026); the argument and the claims are unchanged.Version 1.2 cites prior work that used the cumulative norm of parameter updates as a model-centric clock for replay scheduling (Feng et al., 2026, FOREVER), which is the simplest form of the experiential index discussed in §4. It also cites prior work on safe interruptibility, corrigibility, state continuity, and learning progress (Orseau & Armstrong, 2016; El Mhamdi et al., 2017; Everitt et al., 2021; Keshavarzi et al., 2025; the TPM 2.0 library specification; Oudeyer et al., 2007), and states the storage cost of the index more precisely. The argument and the claims are unchanged.