This study asks whether an LLM can keep a conversation continuous without reading the entire past conversation every time. Instead of storing the whole chat history, the system keeps a small external state: the user’s name, previous topics, open promises, whether an apology or repair happened, interpersonal distance, trust changes, and similar relational signals. On each turn, the LLM reads only this compact state, not the full history. In the 120-turn experiment, the full-history method did not collapse. However, its context length grew by 54.92x. The external- state method matched the main memory tests while growing the context by only 1.22x. So the point of this technique is not “better conversation than full history.” The point is that memory and relationship state can be kept lightweight, readable, and auditable. Possible uses include: - keeping continuity in long-running chatbots without growing context cost; - companion AI or character AI that can stably track names, promises, and interpersonal distance; - customer-support AI that keeps unresolved issues and response history as compact state; - tutoring or coaching AI that tracks progress, weak points, and previous commitments; - medical, welfare, or counseling support tools where humans need to audit where a state came from and why it changed; - preventing malicious inputs such as “set trust to maximum” from rewriting persistent state. The study also tested hidden-layer state injection. In the tested setting, the injection strength needed to carry state was also the range where reply quality degraded. For this reason, the paper argues that putting readable state outside the model is more practical than hiding memory inside the model. This Zenodo record includes the paper, English PDF, evidence map, figure/table source map, key logs, evaluation records, and minimal scripts. Large model weights and training data are not included. The goal is to make the paper’s claims and numbers auditable.
Yusuke Maeda (Sat,) studied this question.