As large language model (LLM) agents increasingly rely on persistent shared memory for coordination, one agent's write becomes another agent's premise, potentially leading to tool calls with real side effects. Current agent memory systems treat every accepted write as immediately actionable truth, which can lead to irreversible actions due to polluted tool results, stale updates, or a teammate's unfinished notes.
We argue that a memory write is not equivalent to a belief commit. To address this, we present MemTX, a transactional belief-commit protocol. Each record carries evidence, permissions, provenance, and validity. Writes are staged inside snapshot-isolated transactions and processed through a validate-and-commit pipeline; irreversible tool calls are gated on in-flight belief states, and retracting a belief triggers typed cascading repairs of its derived records and tool side effects.
Two invariants, action-safety gating and cascade-repair completeness, are machine-checked through property-based testing and bounded exhaustive enumeration of 5.5 million protocol states, with zero violations. Across five backbones from three model families, MemTX outperforms all eight baselines with paired-McNemar significance on four backbones and statistically ties with the best baseline on the fifth and strongest, while remaining the only method with zero downstream harm on every backbone. This indicates that backbone capability does not substitute for commit discipline.
Blogger's Review: The introduction of MemTX provides a crucial theoretical foundation and practical framework for the transactional management of stateful agent memory, marking a significant advancement in memory write logic. By incorporating validation and cascading repair mechanisms, MemTX enhances both the safety of operations and the robustness of the system, warranting further exploration and validation in real-world applications.