Mnemosyne agentic transaction system: validation & repair

Mnemosyne is a runtime that implements Agentic Transaction Processing, a transaction model that treats generated actions as untrusted proposals until they pass deterministic admission under a declared, executable constraint set C. The system and its paper were submitted on 30 Jun 2026 by Edward Y. Chang, Longling Geng and Emily J. Chang.

What is Agentic Transaction Processing (ATP)?

ATP is a two-sided transaction principle that requires runtime admission before proposals become committed state. In one clear formulation the authors write, "a proposal is not truth, and no proposal foresees every disruption." Anything may propose, but only the runtime admits and commits; when an unforeseen disruption strikes the runtime repairs reactively within bounds rather than trusting a fresh proposal.

This means generated workflow actions from LLMs, solvers, or agent teams remain untrusted until they satisfy a deterministic constraint set C. The admission decision and subsequent commit make the correctness of committed state independent of the competence, honesty, or learning of the proposing layer, the paper argues.

How does Mnemosyne enforce validation and repair?

Mnemosyne realizes ATP as a runtime with an append-only transition log, effective-state projection, dependency-safe compensation, and active commitment records. The authors prove four safety properties relative to C: authority separation, serial-equivalent generative admission, evidence-preserving repair, and obligation containment.

Mnemosyne also provides a localized repair protocol, LCRP, and proves a bounded-reactive-repair guarantee for it. The runtime ships with a reproducible artifact that rejects the targeted violations across nine falsification tests while still admitting valid work, and it does so at under 6% projection-and-validation overhead. In addition, the paper reports that bounded local repair edits an order of magnitude fewer operations than global recompute.

Why does the paper claim these safety guarantees?

The safety claims rest on separating proposal generation from deterministic admission and on concrete runtime primitives. The append-only transition log preserves evidence of proposals and transitions. Effective-state projection lets the runtime evaluate proposed transitions against C before commit. Dependency-safe compensation and active commitment records scope repairs so they preserve evidence and contain obligations. Together these primitives underpin the four named safety properties and the bounded-reactive-repair guarantee for LCRP.

The authors support these claims with a reproducible artifact and falsification tests; the artifact rejects the targeted violations in nine tests while admitting valid work. The paper documents the results and the protocol proofs across 36 pages with 24 tables and 6 figures.

Why it matters

Mnemosyne addresses a concrete mismatch: generated actions from agentic systems can be syntactically valid yet stale, infeasible, conflicting, or destructive of the evidence that triggered a repair. By forcing deterministic admission under C and by providing localized, evidence-preserving repair, Mnemosyne shifts correctness from trusting generators to enforcing runtime constraints. That reduces the dependency of committed state on the variable quality of proposing agents and limits harmful global recomputes.

What to watch

Watch for community adoption of the open-source artifact and the falsification tests the authors publish with the paper, and for other systems to report comparable overhead and local-repair savings. A second signal will be independent reproductions of the paper's claim of under 6% projection-and-validation overhead and the reported order-of-magnitude reduction in edits by local repair versus global recompute.

References and provenance The paper, "Mnemosyne: Agentic Transaction Processing for Validating and Repairing AI-generated Workflows," was submitted to arXiv on 30 Jun 2026 by Edward Y. Chang, Longling Geng, and Emily J. Chang. The authors make their implementation and reproducible artifact available as open source in the paper materials.