Mnemosyne: Agentic Transaction Processing for Validating and Repairing AI-generated Workflows

LLMs 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 governing 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; 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) plus a bounded-reactive-repair guarantee (LCRP). A reproducible artifact rejects the targeted violations across nine falsification tests while still admitting valid work, at under 6% overhead, and local repair edits an order of magnitude fewer operations than global recompute. In live-proposer pilots, 80 static plan-entry and mid-execution repair proposals from four heterogeneous LLMs pass the same admission boundary, scored by an external cross-episode harness with zero invalid commits; the gate admits 24 of 40 live repair proposals and rejects 16, four as explicit safety rejections of over-broad rollback. Mnemosyne is open source: https://github.com/eyuchang/Mnemosyne/tree/mnemosyne-atp-postgres-rerun