Mitigating Provenance-Role Collapse in Long-Term Agents via Typed Memory Representation

Problem
This preprint addresses a significant gap in the capability of long-term memory architectures for large language model (LLM) agents, specifically the issue of provenance-role collapse. Current systems typically store historical interactions as unstructured text, leading to source-monitoring errors where agents fail to accurately track the origins and roles of information. This paper highlights the need for a structured approach to memory representation that mitigates these cognitive vulnerabilities.

Method
The authors introduce MemIR, a typed Memory Intermediate Representation designed to enhance source monitoring through a structured memory architecture. MemIR organizes long-term memory into grounded atoms, which compartmentalize raw evidence, retrieval cues, and truth-bearing claims. This architecture enforces a structural constraint on factual authorization, allowing only supported claim atoms to be utilized in decision-making. The method employs multi-route atomic projection and provenance-scoped utilization to convert heterogeneous retrieval hits into claim-centered candidate bundles. This results in a normalized fact interface that facilitates more accurate answer generation. The training and evaluation of MemIR were conducted on two benchmarks: LoCoMo and BEAM-100K, although specific training compute details were not disclosed.

Results
MemIR demonstrates superior performance compared to existing memory baselines across multiple tasks that require source tracking, temporal grounding, and the aggregation of fragmented evidence. On the LoCoMo benchmark, MemIR achieved a 15% improvement in accuracy over the best-performing baseline, while on BEAM-100K, it outperformed the leading model by 20% in tasks involving complex evidence synthesis. These results indicate a significant effect size, underscoring the effectiveness of the proposed memory architecture in addressing the identified issues.

Limitations
The authors acknowledge that while MemIR improves source monitoring, it may introduce additional complexity in memory management and retrieval processes. They also note that the performance gains are primarily observed in tasks that explicitly require source tracking, suggesting that the benefits may not generalize to all LLM applications. Furthermore, the paper does not address the scalability of the architecture in terms of memory size or the computational overhead introduced by the structured representation.

Why it matters
The implications of this work are substantial for the development of more reliable LLM agents capable of maintaining coherent long-term interactions. By addressing provenance-role collapse, MemIR paves the way for enhanced trustworthiness in AI systems that rely on historical data for decision-making. This research could influence future architectures in memory representation, leading to more robust models that can effectively manage and utilize long-term memory in complex tasks. The structured approach proposed by MemIR may also inspire further innovations in cognitive architectures for AI, particularly in applications requiring high levels of accountability and traceability.

Authors: Zhengda Jin, Bingbing Wang, Jing Li, Ruifeng Xu, Min Zhang
Source: arXiv:2605.25869
URL: https://arxiv.org/abs/2605.25869v1