Mnemosyne Introduces Agentic Transaction Processing for AI Workflows

This paper introduces Mnemosyne, a runtime system that implements Agentic Transaction Processing (ATP), a novel transaction model designed to validate and repair AI-generated workflow actions. As large language models (LLMs), solvers, and agent teams increasingly propose actions, there's a risk that these actions, while syntactically valid, might be stale, infeasible, conflicting, or destructive. ATP addresses this by treating all generated actions as untrusted proposals. Under ATP, proposals are only admitted and committed if they pass deterministic validation against a predefined, executable set of constraints. This principle ensures that the correctness of the committed state remains independent of the AI's competence or learning. Mnemosyne features an append-only transition log, effective-state projection, dependency-safe compensation, and active commitment records. The system proves several safety properties and offers a bounded-reactive-repair guarantee, efficiently correcting disruptions locally rather than requiring a global recompute. An open-source artifact demonstrates its ability to reject violations while admitting valid work with minimal overhead.