Get started
Work

LLM Memory, Records, and Continued Use: Content, Provenance, Updates, and Authorization — MRSI Collected Edition

Year: 2026
Type: preprint
Abstract: A correct answer can carry the wrong source identifier. An LLM can faithfully read a stale record or faithfully execute a poor agreement. How can we tell which part of a multi-role system worked, and where an error first entered the chain? This English collected edition brings together four exploratory reports and a technical synthesis. In synthetic two- and three-role tasks, the studies trace what each role received, what it wrote, what was actually saved, and how later calls used it. The first report follows content and provenance through memory updates and revocation. Deterministic copying preserved the targeted shared observation rows, while both methods achieved end-to-end success in 59 of 64 groups, with one rewriting outcome unresolved. The second report shows how a correct joint answer can coexist with a stale private observation, and how a successful repair to one field can be hidden by a failure elsewhere. The third report follows role-specific judgments into joint decisions. Correct individual judgments coexisted with agreement on a lower-scoring proposal, followed by faithful execution. The fourth report traces actual votes in a three-role task: in one readout, a missing vote was reported as an explicit hold even though the group correctly deferred agreement and execution. The concluding synthesis connects six observed failure patterns to seven design principles, a reference architecture, and executable examples. Each report retains its own evaluation units and scope. This record contains the 113-page collected PDF. Standalone reports, readable code, saved experiment records, and verification instructions are available at https://github.com/metaSATOKEN/MRSI.
DOI: 10.5281/zenodo.23971185
Language: English
Open Access status: green
OpenAlex ID: W7225418317
Cites:
Cited by:
Related to: 10