Mnemosyne: Agentic Transaction Processing for Validating and Repairing AI-generated Workflows
06:00 · July 2, 2026 · arXiv cs.AI RSS
LLMs, solvers, and agent teams increasingly generate workflow actions, repairs, and plans, but a generated action may be syntactically valid yet stale, infeasible, conflicting, or destructive of the evidence that triggered a repair. We introduce Agentic Transaction Processing (ATP), a transaction model that treats generated actions as untrusted proposals until they pass deterministic admission under a declared, executable constraint set C. The principle is two-sided: a proposal is not truth, and no proposal foresees every disruption: anything may propose, but only the runtime admits and commits, and when an unforeseen disruption strikes it repairs reactively within bounds rather than trusting a fresh proposal. Relative to C, committed-state correctness becomes independent of the competence, honesty, or learning of the proposing layer. We realize ATP in Mnemosyne, a runtime with an append-only transition log, effective-state projection, dependency-safe compensation, and active commitment records, and prove four safety properties relative to C (authority separation, serial-equivalent generative admission, evidence-preserving repair, and obligation containment) together with a bounded-reactive-repair guarantee for its localized repair protocol (LCRP). A reproducible artifact rejects the targeted violations across nine falsification tests while still admitting valid work, at under 6% projection-and-validation overhead, and bounded local repair edits an order of magnitude fewer operations than global recompute. Mnemosyne is open source: https://github.com/eyuchang/Mnemosyne/tree/arxiv-atp-rq1-rq9b-r8-v2.
Summary
LLMs now routinely draft workflow actions, repairs, and plans, yet a syntactically valid output can still be stale, physically impossible, internally inconsistent, or destructive of prior commitments. Agentic Transaction Processing (ATP) addresses this by treating every generated action as an untrusted proposal that must pass deterministic admission against a declared, executable constraint set C before it can affect committed state. The runtime therefore decouples correctness from the competence or honesty of the proposing layer: anything may suggest a transition, but only the admission gate may commit it.
Mnemosyne implements ATP over an append-only transition log that records every proposal and its outcome. An effective-state projection maintains the current view for constraint checking, while active commitment records track obligations that may later wake and issue repair proposals. When an unforeseen disruption occurs, the localized cascading repair protocol (LCRP) re-plans only the affected fragment and its dependents; the resulting repair re-enters the same deterministic gate, preserving evidence and containing blast radius. Four safety properties—authority separation, serial-equivalent generative admission, evidence-preserving repair, and obligation containment—together with a bounded-reactive-repair guarantee, ensure that committed state remains correct relative to C regardless of how proposals are generated.
Empirical evaluation shows the approach rejects targeted violations across nine falsification tests while admitting valid work, at under 6 % projection-and-validation overhead. Local repair edits an order of magnitude fewer operations than global recompute. In live-proposer pilots, 80 static and mid-execution proposals from four heterogeneous LLMs produced zero invalid commits, and the gate correctly rejected 16 of 40 live repair proposals, including four explicit safety violations involving over-broad rollback. Mnemosyne is released as open source.
Why it matters
This research is highly relevant to the Dutch AI market as it provides a deterministic safety layer for AI agents, aligning perfectly with the EU AI Act's emphasis on transparent, safe, and reliable AI systems. Dutch researchers and enterprises can leverage this open-source framework to build compliant and robust agentic workflows.
