Bounded Numerical Experience Memory with Frozen Language Model Features: Local Correction, Two-Hop Queries, and State Interventions

Abstract

Can a bounded numerical state retain and revise experienced information when language-model weights, auxiliary parameters, interaction history, and native caches cannot carry updates? We investigate this question in a controlled Chinese question-answering setting. The system combines frozen Qwen3.5-0.8B-Base features with learned record routing, reversible character codes, a complete-name span predictor, and a neural extractive output head. Its mutable state occupies 1,095,595 bytes and holds at most seven records. All parameters are fixed during evaluation; only observation writes modify state. In a final, separately generated set of 96 episodes, each containing six facts and one correction, pre- and post-correction fact accuracy is 97.92% and 98.26%, respectively. Two-hop query accuracy is 97.92% before and 96.88% after correction. No incorrect correction allocation is observed in 96 updates, and no false overwrite is observed in 576 initial writes. On a prespecified 128-question intervention subset, normal memory yields 126 correct answers, whereas empty state and disabled reading yield none. Removing semantic features from the extraction head, while preserving retrieval, yields seven correct answers. Cross-process replay reproduces all 1,536 main outputs. These findings support bounded numerical retention, local revision, and state-dependent querying. They do not establish equivalent free generation by the original decoder, long-term open-domain memory, or the formation of new concepts. The study provides an implementation-level mechanism analysis rather than a comparative benchmark claim. English research preprint by baiduman, Wuxi Tianzhiyan Artificial Intelligence. Correspondence: baiduman2020@gmail.com. This controlled mechanism study is not peer reviewed. The companion evidence package includes code, synthetic data, and audit receipts; it contains no model weights. The complete model release is available at https://huggingface.co/duman2020/AD1E-Memory-0.8B-v0.1 (fixed revision b1fa15fb6c2555d1e1b0acefc2e38f94a720e43c). Its published manifest matches all 99 frozen source hashes and the four required auxiliary checkpoint hashes. The manuscript and evidence narrative use CC BY 4.0; existing research source files and linked model release retain Apache-2.0.