From Memory to Causal History: Evidence-Bounded Retrodiction for Persistent Agents
Year: 2026
Type: preprint
Abstract: Persistent AI agents need more than memory. They need a disciplined way to reason about the past without silently giving historical actors information that only became available later. This paper introduces Evidence-Bounded Retrodiction (EBR), a protocol for persistent agents that separates three distinct evidence views: what was contemporaneously recorded, what later evidence supports an actor having access to at the time, and what can be reconstructed retrospectively about the causal history that produced a present state. The central constraint is non-retroactivity: later evidence may improve our understanding of what happened, but it must not rewrite the historical information state against which earlier reasoning and decisions are evaluated. EBR combines provenance-aware temporal memory, actor-access assertions, mechanism-gated candidate histories, contradiction-preserving retrieval, open-set uncertainty, and explicit abstention. Rather than forcing incomplete explanations into a single confidence score, the protocol preserves unresolved alternatives and allows the system to state when the available evidence is insufficient. A deterministic synthetic prototype demonstrates why event time and database record time alone are inadequate proxies for historical epistemic state. In particular, an event may have occurred before a decision but only been discovered afterward, while information available to an actor before the decision may be recorded in a system only later. EBR introduces actor-access evidence to distinguish these cases. The paper also proposes a falsifiable evaluation framework covering temporal leakage, actor-state fidelity, causal-path recovery, contradiction handling, historical-state invariance, and abstention behavior. The contribution is presented as an architectural integration and research protocol for persistent AI agents, not as a new general theory of causality, temporal databases, provenance, or abductive inference. Its narrower aim is to make retrospective reasoning more disciplined: allow an AI to learn from history without rewriting history.
Author: Gill, Ricardo
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7212028289
Cites:
Cited by:
Related to: