Abstract
A personal memory for an AI agent keeps records as nodes of a graph with typed edges and times. When the same thing is written several times as it changes, a search returns old and new records side by side, and the agent may answer from the old one. In this design every record a search returns that has a later record about the same thing carries the id and date of the newest record of its chain, found during the memory's offline maintenance; the search result itself is not reordered, filtered or replaced. The agent can also ask the graph, for a given record, for one of three answers, each a separate operation over fixed edge types: what is current now and what it replaced; how the thing changed over time, with recurrences and a measure of drift; and the different sides held about it, selected by maximal marginal relevance. Every answer is a small subgraph of real records with their dates and ids, no language model writes it, nothing is written back into the graph, and when nothing fits the answer says that the graph does not know.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Prokhorov, Artem, "The Graph Answers: A Personal Memory That Points Every Outdated Record to Its Newest Replacement", Technical Disclosure Commons, (November 05, 2026)
https://www.tdcommons.org/dpubs_series/12772