From Unstructured Recall to Schema-Grounded Memory: Reliable AI Memory via Iterative, Schema-Aware Extraction

Problem
This preprint addresses the limitations of current AI memory systems, which primarily function as retrieval mechanisms for unstructured text. Existing approaches are inadequate for production environments where memory must support operations like exact fact retrieval, state updates, and structured queries. The authors argue that a schema-grounded memory system is necessary to ensure reliability and precision in memory operations, moving beyond simple thematic recall to a more robust system of record.

Method
The authors propose an iterative, schema-aware memory architecture that enhances the write path for memory ingestion. This architecture includes three key components: object detection, field detection, and field-value extraction, all governed by validation gates and local retries. The system employs stateful prompt control to ensure that memory entries adhere to predefined schemas, which dictate what information is retained, ignored, or inferred. This design shifts the focus from the read path—where retrieval is often imprecise—to a more structured write path that ensures data integrity and accuracy.

Results
The proposed memory system, termed xmemory, was evaluated against structured extraction and end-to-end memory benchmarks. On the structured extraction benchmark, xmemory achieved an object-level accuracy of 90.42% and an output accuracy of 62.67%, outperforming all tested frontier structured-output baselines. In the end-to-end memory benchmark, xmemory attained an F1 score of 97.10%, significantly higher than the 80.16%-87.24% range of third-party baselines. Additionally, in application-level tasks, xmemory reached an accuracy of 95.2%, surpassing specialized memory systems and other advanced application harnesses. These results indicate that for memory workloads requiring stable facts and stateful computation, architectural design is more critical than retrieval scale or model strength.

Limitations
The authors acknowledge that their approach may require extensive schema design, which could limit flexibility in dynamic environments. They also note that the performance metrics are contingent on the quality of the schemas defined, which may not generalize across all domains. Additionally, the reliance on a judge-in-the-loop configuration for structured extraction may introduce biases or inconsistencies in evaluation. The paper does not address potential scalability issues when applied to larger datasets or more complex schemas.

Why it matters
This work has significant implications for the development of reliable AI systems that require persistent memory. By establishing a schema-grounded approach, the authors provide a framework that can enhance the accuracy and reliability of memory operations in AI applications. This could lead to advancements in various domains, including conversational agents, automated knowledge management systems, and any application where precise memory retrieval and state management are critical. The findings suggest that future research should focus on optimizing schema design and exploring the integration of this memory architecture into existing AI frameworks.

Authors: Alex Petrov, Alexander Gusak, Denis Mukha, Dima Korolev
Source: arXiv:2604.27906
URL: https://arxiv.org/abs/2604.27906v1