MemTxn: A Transaction Boundary for Source-Supported Updates and Complete-State Recovery in Agent Memory

2026-07-30

AI Summary

This work addresses critical challenges in persistent memory management for large language model agents, including error propagation, unreliable updates, and lack of fault recovery. The paper introduces MemTxn, the first governance layer that enforces transactional boundaries over agent memory. MemTxn ensures reliable memory management through source-validated updates, temporal version resolution, and snapshot-log-driven state recovery, enabling consistent declarative state reconstruction without relying on physical write sets. The system employs an Ordered PatchTest to validate writes, a Temporal Resolver to handle conflicts, and supports full fault recovery. Experimental results demonstrate that MemTxn accepts all valid updates and rejects all hard negative samples in audit tests, achieves exact state recovery on LongMemEval-S and LoCoMo benchmarks, and significantly outperforms the Dense baseline by 17.06–24.07 F1 points on the FactConsolidation task in MemoryAgentBench.

Abstract

Persistent memory lets long-running large language model agents reuse information across sessions and tasks. Yet errors in writable memory can persist and corrupt future behavior. Existing systems improve storage and retrieval, but they do not provide a transaction boundary for reliable updates and recovery. We therefore propose MemTxn, a governance layer outside the answer model. MemTxn verifies whether an update is supported by its source. It also selects the visible version when facts conflict and restores the application-visible state after a fault. The system uses Ordered PatchTest to validate writes, a Temporal Resolver to select versions, and a durable snapshot journal to recover state. On an item-disjoint audit, MemTxn accepts all 60 supported originals and rejects all 179 hard negatives. Under persistent multi-key faults on LongMemEval-S and LoCoMo states, it restores the complete declared active map without knowing the actual physical write set. On MemoryAgentBench FactConsolidation, MemTxn achieves the highest average F1 across all twelve answer-model configurations. It outperforms Dense by 17.06--24.07 points in five representative settings.

Problem

• persistent memory
• transaction boundary
• source-supported updates
• state recovery
• memory consistency

Innovation

• transaction boundary
• source-supported update
• complete-state recovery
• persistent memory
• memory consistency